happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into compilade/test-model-random

This commit is contained in:
Francis Couture-Harpin 2025-07-07 19:53:49 -04:00
parent 7c3f9c226f e1a7059053
commit 48a5eba586
211 changed files with 8853 additions and 7357 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
.devops/tools.sh
View File

@ -1,4 +1,4 @@
#!/bin/bash
#!/usr/bin/env bash
set -e
# Read the first argument into a variable

2
.github/ISSUE_TEMPLATE/010-bug-compilation.yml vendored
View File

@ -40,7 +40,7 @@ body:
attributes:
label: GGML backends
description: Which GGML backends do you know to be affected?
options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
multiple: true
validations:
required: true

2
.github/ISSUE_TEMPLATE/011-bug-results.yml vendored
View File

@ -42,7 +42,7 @@ body:
attributes:
label: GGML backends
description: Which GGML backends do you know to be affected?
options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
multiple: true
validations:
required: true

11
.github/labeler.yml vendored
View File

@ -1,10 +1,4 @@
# https://github.com/actions/labeler
Kompute:
- changed-files:
- any-glob-to-any-file:
- ggml/include/ggml-kompute.h
- ggml/src/ggml-kompute/**
- README-kompute.md
Apple Metal:
- changed-files:
- any-glob-to-any-file:
@ -93,3 +87,8 @@ Ascend NPU:
- ggml/include/ggml-cann.h
- ggml/src/ggml-cann/**
- docs/backend/CANN.md
OpenCL:
- changed-files:
- any-glob-to-any-file:
- ggml/include/ggml-opencl.h
- ggml/src/ggml-opencl/**

30
.github/workflows/build.yml vendored
View File

@ -84,7 +84,8 @@ jobs:
-DCMAKE_BUILD_RPATH="@loader_path" \
-DLLAMA_FATAL_WARNINGS=ON \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DGGML_METAL_EMBED_LIBRARY=OFF \
-DGGML_METAL_SHADER_DEBUG=ON \
-DGGML_RPC=ON
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
@ -664,7 +665,7 @@ jobs:
./build-xcframework.sh
windows-msys2:
runs-on: windows-latest
runs-on: windows-2025
strategy:
fail-fast: false
@ -714,7 +715,7 @@ jobs:
cmake --build build --config ${{ matrix.build }} -j $(nproc)
windows-latest-cmake:
runs-on: windows-latest
runs-on: windows-2025
env:
OPENBLAS_VERSION: 0.3.23
@ -725,17 +726,20 @@ jobs:
matrix:
include:
- build: 'cpu-x64 (static)'
arch: 'x64'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
- build: 'openblas-x64'
arch: 'x64'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
- build: 'vulkan-x64'
arch: 'x64'
defines: '-DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
- build: 'llvm-arm64'
arch: 'arm64'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
- build: 'llvm-arm64-opencl-adreno'
arch: 'arm64'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
# - build: 'kompute-x64'
# defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
steps:
- name: Clone
@ -749,12 +753,6 @@ jobs:
variant: ccache
evict-old-files: 1d
- name: Clone Kompute submodule
id: clone_kompute
if: ${{ matrix.build == 'kompute-x64' }}
run: |
git submodule update --init ggml/src/ggml-kompute/kompute
- name: Download OpenBLAS
id: get_openblas
if: ${{ matrix.build == 'openblas-x64' }}
@ -770,7 +768,7 @@ jobs:
- name: Install Vulkan SDK
id: get_vulkan
if: ${{ matrix.build == 'kompute-x64' || matrix.build == 'vulkan-x64' }}
if: ${{ matrix.build == 'vulkan-x64' }}
run: |
curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
& "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
@ -805,6 +803,8 @@ jobs:
- name: libCURL
id: get_libcurl
uses: ./.github/actions/windows-setup-curl
with:
architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
- name: Build
id: cmake_build
@ -825,7 +825,7 @@ jobs:
- name: Test
id: cmake_test
if: ${{ matrix.build != 'llvm-arm64' && matrix.build != 'llvm-arm64-opencl-adreno' }}
if: ${{ matrix.arch == 'x64' }}
run: |
cd build
ctest -L main -C Release --verbose --timeout 900
@ -930,7 +930,7 @@ jobs:
cmake --build build --config Release
windows-latest-cmake-sycl:
runs-on: windows-latest
runs-on: windows-2022
defaults:
run:
@ -964,7 +964,7 @@ jobs:
windows-latest-cmake-hip:
if: ${{ github.event.inputs.create_release != 'true' }}
runs-on: windows-latest
runs-on: windows-2022
steps:
- name: Clone

22
.github/workflows/release.yml vendored
View File

@ -49,7 +49,8 @@ jobs:
run: |
sysctl -a
cmake -B build \
-DCMAKE_BUILD_RPATH="@loader_path" \
-DCMAKE_INSTALL_RPATH='@loader_path' \
-DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-DLLAMA_FATAL_WARNINGS=ON \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
@ -103,7 +104,8 @@ jobs:
# Metal is disabled due to intermittent failures with Github runners not having a GPU:
# https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
cmake -B build \
-DCMAKE_BUILD_RPATH="@loader_path" \
-DCMAKE_INSTALL_RPATH='@loader_path' \
-DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-DLLAMA_FATAL_WARNINGS=ON \
-DGGML_METAL=OFF \
-DGGML_RPC=ON
@ -160,6 +162,8 @@ jobs:
id: cmake_build
run: |
cmake -B build \
-DCMAKE_INSTALL_RPATH='$ORIGIN' \
-DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-DGGML_BACKEND_DL=ON \
-DGGML_NATIVE=OFF \
-DGGML_CPU_ALL_VARIANTS=ON \
@ -211,6 +215,8 @@ jobs:
id: cmake_build
run: |
cmake -B build \
-DCMAKE_INSTALL_RPATH='$ORIGIN' \
-DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-DGGML_BACKEND_DL=ON \
-DGGML_NATIVE=OFF \
-DGGML_CPU_ALL_VARIANTS=ON \
@ -235,7 +241,7 @@ jobs:
name: llama-bin-ubuntu-vulkan-x64.zip
windows-cpu:
runs-on: windows-latest
runs-on: windows-2025
strategy:
matrix:
@ -271,7 +277,7 @@ jobs:
env:
CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
run: |
call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch }}
call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
cmake -S . -B build -G "Ninja Multi-Config" ^
-D CMAKE_TOOLCHAIN_FILE=cmake/${{ matrix.arch }}-windows-llvm.cmake ^
-DGGML_NATIVE=OFF ^
@ -288,7 +294,7 @@ jobs:
CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
run: |
Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.42.34433\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.44.35112\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
7z a llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
- name: Upload artifacts
@ -298,7 +304,7 @@ jobs:
name: llama-bin-win-cpu-${{ matrix.arch }}.zip
windows:
runs-on: windows-latest
runs-on: windows-2025
env:
OPENBLAS_VERSION: 0.3.23
@ -448,7 +454,7 @@ jobs:
name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
windows-sycl:
runs-on: windows-latest
runs-on: windows-2022
defaults:
run:
@ -520,7 +526,7 @@ jobs:
name: llama-bin-win-sycl-x64.zip
windows-hip:
runs-on: windows-latest
runs-on: windows-2022
strategy:
matrix:

3
.gitmodules vendored
View File

@ -1,3 +0,0 @@
[submodule "kompute"]
path = ggml/src/ggml-kompute/kompute
url = https://github.com/nomic-ai/kompute.git

1
CMakeLists.txt
View File

@ -120,7 +120,6 @@ endfunction()
llama_option_depr(FATAL_ERROR LLAMA_CUBLAS GGML_CUDA)
llama_option_depr(WARNING LLAMA_CUDA GGML_CUDA)
llama_option_depr(WARNING LLAMA_KOMPUTE GGML_KOMPUTE)
llama_option_depr(WARNING LLAMA_METAL GGML_METAL)
llama_option_depr(WARNING LLAMA_METAL_EMBED_LIBRARY GGML_METAL_EMBED_LIBRARY)
llama_option_depr(WARNING LLAMA_NATIVE GGML_NATIVE)

2
build-xcframework.sh
View File

@ -1,4 +1,4 @@
#!/bin/bash
#!/usr/bin/env bash
#
# Options
IOS_MIN_OS_VERSION=16.4

2
ci/run.sh
View File

@ -1,4 +1,4 @@
#!/bin/bash
#!/usr/bin/env bash
#
# sample usage:
#

10
common/arg.cpp
View File

@ -2794,6 +2794,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.ssl_file_cert = value;
}
).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_SSL_CERT_FILE"));
add_opt(common_arg(
{"--chat-template-kwargs"}, "STRING",
string_format("sets additional params for the json template parser"),
[](common_params & params, const std::string & value) {
auto parsed = json::parse(value);
for (const auto & item : parsed.items()) {
params.default_template_kwargs[item.key()] = item.value().dump();
}
}
).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_CHAT_TEMPLATE_KWARGS"));
add_opt(common_arg(
{"-to", "--timeout"}, "N",
string_format("server read/write timeout in seconds (default: %d)", params.timeout_read),

57
common/chat.cpp
View File

@ -17,6 +17,8 @@
#include <string>
#include <vector>
using json = nlohmann::ordered_json;
static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
auto time = std::chrono::system_clock::to_time_t(now);
auto local_time = *std::localtime(&time);
@ -140,6 +142,7 @@ struct templates_params {
bool add_generation_prompt = true;
bool enable_thinking = true;
std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
json extra_context;
};
common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) {
@ -720,16 +723,23 @@ static void foreach_function(const json & tools, const std::function<void(const
static std::string apply(
const common_chat_template & tmpl,
const nlohmann::ordered_json & messages,
const nlohmann::ordered_json & tools,
bool add_generation_prompt,
const nlohmann::ordered_json & extra_context = nlohmann::ordered_json())
const struct templates_params & inputs,
const std::optional<json> & messages_override = std::nullopt,
const std::optional<json> & tools_override = std::nullopt,
const std::optional<json> & additional_context = std::nullopt)
{
minja::chat_template_inputs tmpl_inputs;
tmpl_inputs.messages = messages;
tmpl_inputs.tools = tools;
tmpl_inputs.add_generation_prompt = add_generation_prompt;
tmpl_inputs.extra_context = extra_context;
tmpl_inputs.messages = messages_override ? *messages_override : inputs.messages;
if (tools_override) {
tmpl_inputs.tools = *tools_override;
} else {
tmpl_inputs.tools = inputs.tools.empty() ? json() : inputs.tools;
}
tmpl_inputs.add_generation_prompt = inputs.add_generation_prompt;
tmpl_inputs.extra_context = inputs.extra_context;
if (additional_context) {
tmpl_inputs.extra_context.merge_patch(*additional_context);
}
// TODO: add flag to control date/time, if only for testing purposes.
// tmpl_inputs.now = std::chrono::system_clock::now();
@ -828,7 +838,7 @@ static common_chat_params common_chat_params_init_generic(const common_chat_temp
inputs.messages,
"Respond in JSON format, either with `tool_call` (a request to call tools) or with `response` reply to the user's request");
data.prompt = apply(tmpl, tweaked_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.prompt = apply(tmpl, inputs, /* messages_override= */ tweaked_messages);
data.format = COMMON_CHAT_FORMAT_GENERIC;
return data;
}
@ -904,7 +914,7 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
data.preserved_tokens = {
"[TOOL_CALLS]",
};
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.prompt = apply(tmpl, inputs);
data.format = COMMON_CHAT_FORMAT_MISTRAL_NEMO;
return data;
}
@ -934,7 +944,7 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_
adjusted_messages.push_back(msg);
}
}
data.prompt = apply(tmpl, adjusted_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {});
data.prompt = apply(tmpl, inputs, /* messages_override= */ adjusted_messages);
data.format = COMMON_CHAT_FORMAT_COMMAND_R7B;
if (string_ends_with(data.prompt, "<|START_THINKING|>")) {
if (!inputs.enable_thinking) {
@ -1122,7 +1132,7 @@ static common_chat_params common_chat_params_init_llama_3_x(const common_chat_te
} else {
data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
}
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {
data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, json {
{"date_string", format_time(inputs.now, "%d %b %Y")},
{"tools_in_user_message", false},
{"builtin_tools", builtin_tools.empty() ? json() : builtin_tools},
@ -1187,7 +1197,7 @@ static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool w
static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_template & tmpl, const struct templates_params & inputs) {
common_chat_params data;
auto prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
auto prompt = apply(tmpl, inputs);
// Hacks to fix the official (broken) prompt.
// It is advisable to use --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja instead,
@ -1282,7 +1292,7 @@ static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) {
LOG_DBG("%s\n", __func__);
common_chat_params data;
data.prompt = apply(tmpl, inputs.messages, /* tools= */ nullptr, inputs.add_generation_prompt, {
data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ json(), json {
{"datetime", format_time(inputs.now, "%b %d %Y %H:%M:%S GMT")},
{"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
});
@ -1338,7 +1348,7 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
// Using ">>>f1\n", ">>>f2\n"... as trigger words for the grammar
// If the function is python, we also allow raw python code (if the line after `python\n` doesn't start w/ opening `{`), which the model seems to prefer for multiline code.
common_chat_params data;
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.prompt = apply(tmpl, inputs);
data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2;
if (inputs.tools.is_array() && !inputs.tools.empty()) {
data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
@ -1465,7 +1475,7 @@ static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(con
data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
}
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.prompt = apply(tmpl, inputs);
// TODO: if (has_raw_python)
return data;
}
@ -1498,14 +1508,15 @@ static void common_chat_parse_functionary_v3_1_llama_3_1(common_chat_msg_parser
static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat_template & tmpl, const struct templates_params & inputs) {
common_chat_params data;
json additional_context = {
json extra_context = json {
{"enable_thinking", inputs.enable_thinking},
};
extra_context.update(inputs.extra_context);
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, additional_context);
data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, extra_context);
data.format = COMMON_CHAT_FORMAT_HERMES_2_PRO;
if (string_ends_with(data.prompt, "<think>\n")) {
if (!inputs.enable_thinking) {
if (!extra_context["enable_thinking"]) {
data.prompt += "</think>";
} else {
data.thinking_forced_open = true;
@ -1691,7 +1702,7 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {
common_chat_params data;
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.prompt = apply(tmpl, inputs);
data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
data.grammar_lazy = false;
if (!inputs.json_schema.is_null()) {
@ -1722,6 +1733,12 @@ static common_chat_params common_chat_templates_apply_jinja(
params.enable_thinking = inputs.enable_thinking;
params.grammar = inputs.grammar;
params.now = inputs.now;
params.extra_context = json::object();
for (auto el : inputs.chat_template_kwargs) {
params.extra_context[el.first] = json::parse(el.second);
}
if (!inputs.json_schema.empty()) {
params.json_schema = json::parse(inputs.json_schema);
}

2
common/chat.h
View File

@ -7,6 +7,7 @@
#include <chrono>
#include <string>
#include <vector>
#include <map>
struct common_chat_templates;
@ -125,6 +126,7 @@ struct common_chat_templates_inputs {
common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE;
bool enable_thinking = true;
std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
std::map<std::string, std::string> chat_template_kwargs;
};
struct common_chat_params {

3
common/common.h
View File

@ -8,6 +8,7 @@
#include <string>
#include <string_view>
#include <vector>
#include <map>
#include <sstream>
#ifdef _WIN32
@ -381,6 +382,8 @@ struct common_params {
std::string ssl_file_key = ""; // NOLINT
std::string ssl_file_cert = ""; // NOLINT
std::map<std::string, std::string> default_template_kwargs;
// "advanced" endpoints are disabled by default for better security
bool webui = true;
bool endpoint_slots = false;

171
convert_hf_to_gguf.py
View File

@ -936,7 +936,11 @@ class TextModel(ModelBase):
scores: list[float] = [-10000.0] * vocab_size
toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
for token_id in range(vocab_size):
for token_id in range(tokenizer.vocab_size()):
if token_id >= vocab_size:
logger.warning(f'ignore tokens from {token_id}: id is out of range, max={vocab_size - 1}')
break
piece = tokenizer.IdToPiece(token_id)
text = piece.encode("utf-8")
score = tokenizer.GetScore(token_id)
@ -951,10 +955,6 @@ class TextModel(ModelBase):
elif tokenizer.IsByte(token_id):
toktype = SentencePieceTokenTypes.BYTE
if token_id >= vocab_size:
logger.warning(f'ignore tokens from {token_id}: id is out of range, max={vocab_size - 1}')
break
tokens[token_id] = text
scores[token_id] = score
toktypes[token_id] = toktype
@ -2743,6 +2743,52 @@ class Qwen2Model(TextModel):
yield from super().modify_tensors(data_torch, name, bid)
@ModelBase.register("Ernie4_5_ForCausalLM")
class Ernie4_5Model(TextModel):
model_arch = gguf.MODEL_ARCH.ERNIE4_5
def set_vocab(self):
self._set_vocab_sentencepiece()
def set_gguf_parameters(self):
super().set_gguf_parameters()
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
num_heads = self.hparams["num_attention_heads"]
num_kv_heads = self.hparams["num_key_value_heads"]
head_dim = self.hparams["head_dim"]
if "ernie." in name:
name = name.replace("ernie.", "model.")
# split the qkv weights
# qkv_proj shape: [(num_heads + 2 * num_kv_heads) * head_dim, hidden_size]
if "qkv_proj" in name:
name_q = name.replace("qkv_proj.weight", "q_proj.weight")
name_k = name.replace("qkv_proj.weight", "k_proj.weight")
name_v = name.replace("qkv_proj.weight", "v_proj.weight")
total_q_dim = num_heads * head_dim
total_k_dim = num_kv_heads * head_dim
total_v_dim = num_kv_heads * head_dim
q_proj_weight, k_proj_weight, v_proj_weight = data_torch.split([total_q_dim, total_k_dim, total_v_dim], dim=0)
return [
(self.map_tensor_name(name_q), q_proj_weight),
(self.map_tensor_name(name_k), k_proj_weight),
(self.map_tensor_name(name_v), v_proj_weight)
]
# split the up_gate_proj into gate and up
# up_gate_proj shape: [2 * intermediate_size, hidden_size]
if "up_gate_proj" in name:
name_up = name.replace("up_gate_proj.weight", "up_proj.weight")
name_gate = name.replace("up_gate_proj.weight", "gate_proj.weight")
dim_half = data_torch.shape[0] // 2
gate_proj_weight, up_proj_weight = data_torch.split(dim_half, dim=0)
return [
(self.map_tensor_name(name_gate), gate_proj_weight),
(self.map_tensor_name(name_up), up_proj_weight)
]
return [(self.map_tensor_name(name), data_torch)]
@ModelBase.register(
"Qwen2VLModel",
"Qwen2VLForConditionalGeneration",
@ -4362,9 +4408,6 @@ class Gemma3NModel(Gemma3Model):
]
def set_vocab(self):
with open(self.dir_model / "chat_template.jinja") as f:
# quick hack to make sure chat template is added
self.gguf_writer.add_chat_template(f.read())
super().set_vocab()
def set_gguf_parameters(self):
@ -4735,6 +4778,14 @@ class ARwkv7Model(Rwkv7Model):
class MambaModel(TextModel):
model_arch = gguf.MODEL_ARCH.MAMBA
def __init__(self, dir_model: Path, *args, **kwargs):
# Avoid using AutoConfig for hparams
hparams = kwargs.pop("hparams", None)
if hparams is None:
with open(dir_model / "config.json", "r", encoding="utf-8") as f:
hparams = json.load(f)
super().__init__(dir_model, *args, hparams=hparams, **kwargs)
def set_vocab(self):
vocab_size = self.hparams["vocab_size"]
# Round vocab size to next multiple of 8
@ -4809,6 +4860,100 @@ class MambaModel(TextModel):
return [(new_name, data_torch)]
@ModelBase.register("Mamba2ForCausalLM")
class Mamba2Model(TextModel):
model_arch = gguf.MODEL_ARCH.MAMBA2
def __init__(self, dir_model: Path, *args, **kwargs):
# Avoid using AutoConfig for hparams
# It wrongly assumes all Mamba2 models are Mamba-Codestral-7B-v0.1
hparams = kwargs.pop("hparams", None)
if hparams is None:
with open(dir_model / "config.json", "r", encoding="utf-8") as f:
hparams = json.load(f)
super().__init__(dir_model, *args, hparams=hparams, **kwargs)
def set_vocab(self):
vocab_size = self.hparams["vocab_size"]
# Round vocab size to next multiple of 16
pad_vocab = self.hparams.get("pad_vocab_size_multiple", 16)
# pad using ceiling division
# ref: https://stackoverflow.com/a/17511341/22827863
vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
self.hparams["vocab_size"] = vocab_size
if (self.dir_model / "tokenizer.model").is_file():
self._set_vocab_sentencepiece()
elif (self.dir_model / "tokenizer.model.v3").is_file():
# mamba-codestral
raise NotImplementedError(f"Please rename {self.dir_model / 'tokenizer.model.v3'} to {self.dir_model / 'tokenizer.model'}")
elif (self.dir_model / "tokenizer.json").is_file():
self._set_vocab_gpt2()
else:
# Use the GPT-NeoX tokenizer when no tokenizer files are present
self._set_vocab_builtin("gpt-neox", vocab_size)
def set_gguf_parameters(self):
d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
d_conv = self.find_hparam(["conv_kernel", "d_conv"], optional=True) or 4
d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
d_state = self.find_hparam(["state_size", "d_state"], optional=True) or 128
head_dim = self.find_hparam(["head_dim"], optional=True) or 64
n_group = self.find_hparam(["n_groups"], optional=True) or 1
rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
# Fail early for models which don't have a block expansion factor of 2
# TODO: does this really matter?
assert d_inner == 2 * d_model
assert d_inner % head_dim == 0
self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default
self.gguf_writer.add_embedding_length(d_model)
self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading
self.gguf_writer.add_head_count(0) # unused, but seemingly required when loading
self.gguf_writer.add_block_count(self.block_count)
self.gguf_writer.add_ssm_conv_kernel(d_conv)
self.gguf_writer.add_ssm_inner_size(d_inner)
self.gguf_writer.add_ssm_state_size(d_state)
self.gguf_writer.add_ssm_time_step_rank(d_inner // head_dim)
self.gguf_writer.add_ssm_group_count(n_group)
self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
self.gguf_writer.add_file_type(self.ftype)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
if name.startswith("model.backbone") or name.startswith("model.lm_head"):
# map Mamba-Codestral-7B-v0.1 tensor names to the names used by Mamba-2
name = name.removeprefix("model.")
if name.endswith(".dt_bias"):
name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
new_name = self.map_tensor_name(name)
if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_CONV1D, bid):
data_torch = data_torch.squeeze()
elif any(self.match_model_tensor_name(new_name, t, bid, suffix="") for t in [
gguf.MODEL_TENSOR.SSM_A,
gguf.MODEL_TENSOR.SSM_D,
]):
# unsqueeze A to use similar shape semantics as Mamba-1
# (D is also unsqueezed, but for more straightforward broadcast internally)
data_torch = data_torch.reshape((*data_torch.shape, 1))
elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_NORM, bid):
d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
n_group = self.hparams.get("n_groups", 1)
data_torch = data_torch.reshape((n_group, d_inner // n_group))
if name.endswith(".A_log"):
logger.debug("A_log --> A ==> " + new_name)
data_torch = -torch.exp(data_torch)
yield (new_name, data_torch)
@ModelBase.register("CohereForCausalLM")
class CommandR2Model(TextModel):
model_arch = gguf.MODEL_ARCH.COMMAND_R
@ -6569,12 +6714,20 @@ def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> st
# maybe we should fallback to text model's arch in that case, since not many models have both
text_config = hparams.get("text_config", {})
vision_config = hparams.get("vision_config", {})
arch = hparams["architectures"][0]
arch = None
if (arches := hparams.get("architectures")) is not None and len(arches) > 0:
arch = arches[0]
elif "ssm_cfg" in hparams:
# For non-hf Mamba and Mamba2 models
arch = hparams["ssm_cfg"].get("layer", "Mamba") + "ForCausalLM"
# if "architectures" is found in the sub-config, use that instead
if model_type == ModelType.TEXT and text_config.get("architectures") is not None:
arch = text_config["architectures"][0]
elif model_type == ModelType.MMPROJ and vision_config.get("architectures") is not None:
arch = vision_config["architectures"][0]
if arch is None:
raise ValueError("Failed to detect model architecture")
return arch

3
docs/docker.md
View File

@ -25,6 +25,9 @@ Additionally, there the following images, similar to the above:
- `ghcr.io/ggml-org/llama.cpp:full-intel`: Same as `full` but compiled with SYCL support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:light-intel`: Same as `light` but compiled with SYCL support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:server-intel`: Same as `server` but compiled with SYCL support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:full-vulkan`: Same as `full` but compiled with Vulkan support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:light-vulkan`: Same as `light` but compiled with Vulkan support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:server-vulkan`: Same as `server` but compiled with Vulkan support. (platforms: `linux/amd64`)
The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA, ROCm or MUSA library, you'll need to build the images locally for now).

2
examples/Miku.sh
View File

@ -1,4 +1,4 @@
#!/bin/bash
#!/usr/bin/env bash
set -e
AI_NAME="${AI_NAME:-Miku}"

2
examples/chat-13B.sh
View File

@ -1,4 +1,4 @@
#!/bin/bash
#!/usr/bin/env bash
set -e

2
examples/chat-persistent.sh
View File

@ -1,4 +1,4 @@
#!/bin/bash
#!/usr/bin/env bash
set -euo pipefail

2
examples/chat-vicuna.sh
View File

@ -1,4 +1,4 @@
#!/bin/bash
#!/usr/bin/env bash
set -e

2
examples/chat.sh
View File

@ -1,4 +1,4 @@
#!/bin/bash
#!/usr/bin/env bash
#
# Temporary script - will be removed in the future

7
examples/eval-callback/eval-callback.cpp
View File

@ -55,6 +55,8 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
} else if (type == GGML_TYPE_F32) {
v = *(float *) &data[i];
} else if (type == GGML_TYPE_I64) {
v = (float) *(int64_t *) &data[i];
} else if (type == GGML_TYPE_I32) {
v = (float) *(int32_t *) &data[i];
} else if (type == GGML_TYPE_I16) {
@ -134,6 +136,11 @@ static bool run(llama_context * ctx, const common_params & params) {
std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
if (tokens.empty()) {
LOG_ERR("%s : there are not input tokens to process - (try to provide a prompt with '-p')\n", __func__);
return false;
}
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
LOG_ERR("%s : failed to eval\n", __func__);
return false;

2
examples/jeopardy/jeopardy.sh
View File

@ -1,4 +1,4 @@
#!/bin/bash
#!/usr/bin/env bash
set -e
MODEL=./models/ggml-vicuna-13b-1.1-q4_0.bin

2
examples/reason-act.sh
View File

@ -1,4 +1,4 @@
#!/bin/bash
#!/usr/bin/env bash
cd `dirname $0`
cd ..

2
examples/server-llama2-13B.sh
View File

@ -1,4 +1,4 @@
#!/bin/bash
#!/usr/bin/env bash
set -e

7
examples/simple-chat/simple-chat.cpp
View File

@ -113,15 +113,16 @@ int main(int argc, char ** argv) {
while (true) {
// check if we have enough space in the context to evaluate this batch
int n_ctx = llama_n_ctx(ctx);
int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx), 0);
int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) + 1;
if (n_ctx_used + batch.n_tokens > n_ctx) {
printf("\033[0m\n");
fprintf(stderr, "context size exceeded\n");
exit(0);
}
if (llama_decode(ctx, batch)) {
GGML_ABORT("failed to decode\n");
int ret = llama_decode(ctx, batch);
if (ret != 0) {
GGML_ABORT("failed to decode, ret = %d\n", ret);
}
// sample the next token

2
examples/sycl/build.sh
View File

@ -1,4 +1,4 @@
#!/usr/bin/env bash
# MIT license
# Copyright (C) 2024 Intel Corporation
# SPDX-License-Identifier: MIT

2
examples/sycl/run-llama2.sh
View File

@ -1,4 +1,4 @@
#!/bin/bash
#!/usr/bin/env bash
# MIT license
# Copyright (C) 2024 Intel Corporation

2
examples/sycl/run-llama3.sh
View File

@ -1,4 +1,4 @@
#!/bin/bash
#!/usr/bin/env bash
# MIT license
# Copyright (C) 2025 Intel Corporation

2
examples/ts-type-to-grammar.sh
View File

@ -1,4 +1,4 @@
#!/bin/bash
#!/usr/bin/env bash
#
# ./examples/ts-type-to-grammar.sh "{a:string,b:string,c?:string}"
# python examples/json_schema_to_grammar.py https://json.schemastore.org/tsconfig.json

9
ggml/CMakeLists.txt
View File

@ -181,7 +181,6 @@ option(GGML_VULKAN_MEMORY_DEBUG "ggml: enable Vulkan memory debug ou
option(GGML_VULKAN_SHADER_DEBUG_INFO "ggml: enable Vulkan shader debug info" OFF)
option(GGML_VULKAN_VALIDATE "ggml: enable Vulkan validation" OFF)
option(GGML_VULKAN_RUN_TESTS "ggml: run Vulkan tests" OFF)
option(GGML_KOMPUTE "ggml: use Kompute" OFF)
option(GGML_METAL "ggml: use Metal" ${GGML_METAL_DEFAULT})
option(GGML_METAL_USE_BF16 "ggml: use bfloat if available" OFF)
option(GGML_METAL_NDEBUG "ggml: disable Metal debugging" OFF)
@ -266,7 +265,6 @@ set(GGML_PUBLIC_HEADERS
include/ggml-cann.h
include/ggml-cpp.h
include/ggml-cuda.h
include/ggml-kompute.h
include/ggml-opt.h
include/ggml-metal.h
include/ggml-rpc.h
@ -360,6 +358,13 @@ write_basic_package_version_file(
VERSION ${GGML_INSTALL_VERSION}
COMPATIBILITY SameMajorVersion)
target_compile_definitions(ggml-base PRIVATE
GGML_VERSION="${GGML_INSTALL_VERSION}"
GGML_COMMIT="${GGML_BUILD_COMMIT}"
)
message(STATUS "ggml version: ${GGML_INSTALL_VERSION}")
message(STATUS "ggml commit: ${GGML_BUILD_COMMIT}")
install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml)

2
ggml/include/ggml-backend.h
View File

@ -339,7 +339,7 @@ extern "C" {
typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
// Compare the output of two backends
GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data, struct ggml_tensor * test_node);
// Tensor initialization
GGML_API enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);

1
ggml/include/ggml-cpu.h
View File

@ -134,6 +134,7 @@ extern "C" {
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
GGML_BACKEND_API void ggml_cpu_fp32_to_fp32(const float *, float *, int64_t);
GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t);
GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t);
GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t);

50
ggml/include/ggml-kompute.h
View File

@ -1,50 +0,0 @@
#pragma once
#include "ggml.h"
#include "ggml-backend.h"
#include <stdbool.h>
#include <stddef.h>
#include <stdint.h>
#ifdef __cplusplus
extern "C" {
#endif
#define GGML_KOMPUTE_MAX_DEVICES 16
struct ggml_vk_device {
int index;
int type; // same as VkPhysicalDeviceType
size_t heapSize;
const char * name;
const char * vendor;
int subgroupSize;
uint64_t bufferAlignment;
uint64_t maxAlloc;
};
struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
bool ggml_vk_has_vulkan(void);
bool ggml_vk_has_device(void);
struct ggml_vk_device ggml_vk_current_device(void);
//
// backend API
//
// forward declaration
typedef struct ggml_backend * ggml_backend_t;
GGML_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
#ifdef __cplusplus
}
#endif

189
ggml/include/ggml.h
View File

@ -314,6 +314,13 @@
extern "C" {
#endif
// Function type used in fatal error callbacks
typedef void (*ggml_abort_callback_t)(const char * error_message);
// Set the abort callback (passing null will restore original abort functionality: printing a message to stdout)
// Returns the old callback for chaining
GGML_API ggml_abort_callback_t ggml_set_abort_callback(ggml_abort_callback_t callback);
GGML_NORETURN GGML_ATTRIBUTE_FORMAT(3, 4)
GGML_API void ggml_abort(const char * file, int line, const char * fmt, ...);
@ -470,6 +477,7 @@ extern "C" {
GGML_OP_TRANSPOSE,
GGML_OP_GET_ROWS,
GGML_OP_GET_ROWS_BACK,
GGML_OP_SET_ROWS,
GGML_OP_DIAG,
GGML_OP_DIAG_MASK_INF,
GGML_OP_DIAG_MASK_ZERO,
@ -481,6 +489,7 @@ extern "C" {
GGML_OP_CONV_TRANSPOSE_1D,
GGML_OP_IM2COL,
GGML_OP_IM2COL_BACK,
GGML_OP_CONV_2D,
GGML_OP_CONV_2D_DW,
GGML_OP_CONV_TRANSPOSE_2D,
GGML_OP_POOL_1D,
@ -519,6 +528,8 @@ extern "C" {
GGML_OP_CROSS_ENTROPY_LOSS_BACK,
GGML_OP_OPT_STEP_ADAMW,
GGML_OP_GLU,
GGML_OP_COUNT,
};
@ -542,6 +553,16 @@ extern "C" {
GGML_UNARY_OP_COUNT,
};
enum ggml_glu_op {
GGML_GLU_OP_REGLU,
GGML_GLU_OP_GEGLU,
GGML_GLU_OP_SWIGLU,
GGML_GLU_OP_GEGLU_ERF,
GGML_GLU_OP_GEGLU_QUICK,
GGML_GLU_OP_COUNT,
};
enum ggml_object_type {
GGML_OBJECT_TYPE_TENSOR,
GGML_OBJECT_TYPE_GRAPH,
@ -627,6 +648,9 @@ extern "C" {
// misc
GGML_API const char * ggml_version(void);
GGML_API const char * ggml_commit(void);
GGML_API void ggml_time_init(void); // call this once at the beginning of the program
GGML_API int64_t ggml_time_ms(void);
GGML_API int64_t ggml_time_us(void);
@ -657,6 +681,7 @@ extern "C" {
GGML_API const char * ggml_op_symbol(enum ggml_op op);
GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op);
GGML_API const char * ggml_glu_op_name(enum ggml_glu_op op);
GGML_API const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
GGML_API size_t ggml_element_size(const struct ggml_tensor * tensor);
@ -687,6 +712,9 @@ extern "C" {
// true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
GGML_API bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor);
// true if the elements in dimension 0 are contiguous, or there is just 1 block of elements
GGML_API bool ggml_is_contiguous_rows(const struct ggml_tensor * tensor);
GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
@ -758,6 +786,7 @@ extern "C" {
GGML_API void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
GGML_API enum ggml_glu_op ggml_get_glu_op(const struct ggml_tensor * tensor);
GGML_API void * ggml_get_data (const struct ggml_tensor * tensor);
GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
@ -1086,6 +1115,89 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a);
// gated linear unit ops
// A: n columns, r rows,
// result is n / 2 columns, r rows,
// expects gate in second half of row, unless swapped is true
GGML_API struct ggml_tensor * ggml_glu(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_glu_op op,
bool swapped);
GGML_API struct ggml_tensor * ggml_reglu(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_reglu_swapped(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_geglu(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_geglu_swapped(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_swiglu(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_swiglu_swapped(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_geglu_erf(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_geglu_erf_swapped(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_geglu_quick(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_geglu_quick_swapped(
struct ggml_context * ctx,
struct ggml_tensor * a);
// A: n columns, r rows,
// B: n columns, r rows,
GGML_API struct ggml_tensor * ggml_glu_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
enum ggml_glu_op op);
GGML_API struct ggml_tensor * ggml_reglu_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_geglu_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_swiglu_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_geglu_erf_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_geglu_quick_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// normalize along rows
GGML_API struct ggml_tensor * ggml_norm(
struct ggml_context * ctx,
@ -1375,6 +1487,23 @@ extern "C" {
struct ggml_tensor * b, // row indices
struct ggml_tensor * c); // data for ggml_get_rows, only used for its shape
// a TD [n_embd, ne1, ne2, ne3]
// b TS [n_embd, n_rows, ne02, ne03] | ne02 == ne2, ne03 == ne3
// c I64 [n_rows, ne11, ne12, 1] | c[i] in [0, ne1)
//
// undefined behavior if destination rows overlap
//
// broadcast:
// ne2 % ne11 == 0
// ne3 % ne12 == 0
//
// return view(a)
GGML_API struct ggml_tensor * ggml_set_rows(
struct ggml_context * ctx,
struct ggml_tensor * a, // destination
struct ggml_tensor * b, // source
struct ggml_tensor * c); // row indices
GGML_API struct ggml_tensor * ggml_diag(
struct ggml_context * ctx,
struct ggml_tensor * a);
@ -1412,8 +1541,14 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a);
// a [ne0, ne01, ne02, ne03]
// mask [ne0, ne11, ne12, ne13] | ne11 >= ne01, F16 or F32, optional
//
// broadcast:
// ne02 % ne12 == 0
// ne03 % ne13 == 0
//
// fused soft_max(a*scale + mask*(ALiBi slope))
// mask is optional
// max_bias = 0.0f for no ALiBi
GGML_API struct ggml_tensor * ggml_soft_max_ext(
struct ggml_context * ctx,
@ -1723,6 +1858,17 @@ extern "C" {
struct ggml_tensor * b,
int stride);
GGML_API struct ggml_tensor * ggml_conv_2d_direct(
struct ggml_context * ctx,
struct ggml_tensor * a, // convolution kernel [KW, KH, IC, OC]
struct ggml_tensor * b, // input data [W, H, C, N]
int s0, // stride dimension 0
int s1, // stride dimension 1
int p0, // padding dimension 0
int p1, // padding dimension 1
int d0, // dilation dimension 0
int d1); // dilation dimension 1
enum ggml_op_pool {
GGML_OP_POOL_MAX,
GGML_OP_POOL_AVG,
@ -1765,6 +1911,12 @@ extern "C" {
enum ggml_scale_mode {
GGML_SCALE_MODE_NEAREST = 0,
GGML_SCALE_MODE_BILINEAR = 1,
GGML_SCALE_MODE_COUNT
};
enum ggml_scale_flag {
GGML_SCALE_FLAG_ALIGN_CORNERS = (1 << 8)
};
// interpolate
@ -1777,14 +1929,26 @@ extern "C" {
// interpolate
// interpolate scale to specified dimensions
GGML_API struct ggml_tensor * ggml_upscale_ext(
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_upscale_ext(
struct ggml_context * ctx,
struct ggml_tensor * a,
int ne0,
int ne1,
int ne2,
int ne3,
enum ggml_scale_mode mode);
enum ggml_scale_mode mode),
"use ggml_interpolate instead");
// Up- or downsamples the input to the specified size.
// 2D scale modes (eg. bilinear) are applied to the first two dimensions.
GGML_API struct ggml_tensor * ggml_interpolate(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
int64_t ne1,
int64_t ne2,
int64_t ne3,
uint32_t mode); // ggml_scale_mode [ | ggml_scale_flag...]
// pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
GGML_API struct ggml_tensor * ggml_pad(
@ -1847,11 +2011,17 @@ extern "C" {
#define GGML_KQ_MASK_PAD 64
// q: [n_embd_k, n_batch, n_head, 1]
// k: [n_embd_k, n_kv, n_head_kv, 1]
// v: [n_embd_v, n_kv, n_head_kv, 1] !! not transposed !!
// mask: [n_kv, n_batch_pad, 1, 1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
// res: [n_embd_v, n_head, n_batch, 1] !! permuted !!
// q: [n_embd_k, n_batch, n_head, ne3 ]
// k: [n_embd_k, n_kv, n_head_kv, ne3 ]
// v: [n_embd_v, n_kv, n_head_kv, ne3 ] !! not transposed !!
// mask: [n_kv, n_batch_pad, ne32, ne33] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
// res: [n_embd_v, n_head, n_batch, ne3 ] !! permuted !!
//
// broadcast:
// n_head % n_head_kv == 0
// n_head % ne32 == 0
// ne3 % ne33 == 0
//
GGML_API struct ggml_tensor * ggml_flash_attn_ext(
struct ggml_context * ctx,
struct ggml_tensor * q,
@ -1890,7 +2060,8 @@ extern "C" {
struct ggml_tensor * dt,
struct ggml_tensor * A,
struct ggml_tensor * B,
struct ggml_tensor * C);
struct ggml_tensor * C,
struct ggml_tensor * ids);
// partition into non-overlapping windows with padding if needed
// example:

1
ggml/src/CMakeLists.txt
View File

@ -365,7 +365,6 @@ ggml_add_backend(BLAS)
ggml_add_backend(CANN)
ggml_add_backend(CUDA)
ggml_add_backend(HIP)
ggml_add_backend(Kompute)
ggml_add_backend(METAL)
ggml_add_backend(MUSA)
ggml_add_backend(RPC)

8
ggml/src/ggml-backend-reg.cpp
View File

@ -61,10 +61,6 @@
#include "ggml-cann.h"
#endif
#ifdef GGML_USE_KOMPUTE
#include "ggml-kompute.h"
#endif
// disable C++17 deprecation warning for std::codecvt_utf8
#if defined(__clang__)
# pragma clang diagnostic push
@ -189,9 +185,6 @@ struct ggml_backend_registry {
#ifdef GGML_USE_RPC
register_backend(ggml_backend_rpc_reg());
#endif
#ifdef GGML_USE_KOMPUTE
register_backend(ggml_backend_kompute_reg());
#endif
#ifdef GGML_USE_CPU
register_backend(ggml_backend_cpu_reg());
#endif
@ -575,7 +568,6 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
ggml_backend_load_best("cann", silent, dir_path);
ggml_backend_load_best("cuda", silent, dir_path);
ggml_backend_load_best("hip", silent, dir_path);
ggml_backend_load_best("kompute", silent, dir_path);
ggml_backend_load_best("metal", silent, dir_path);
ggml_backend_load_best("rpc", silent, dir_path);
ggml_backend_load_best("sycl", silent, dir_path);

58
ggml/src/ggml-backend.cpp
View File

@ -817,8 +817,9 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
}
if (sched->debug > 1) {
ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s] use=%d:", i, ggml_op_name(node->op), node->name,
fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node),
graph->use_counts[ggml_hash_find(&graph->visited_hash_set, node)]);
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (src == NULL) {
@ -1826,7 +1827,7 @@ void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy) {
ggml_free(copy.ctx_unallocated);
}
bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data) {
bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data, struct ggml_tensor * test_node) {
struct ggml_backend_graph_copy copy = ggml_backend_graph_copy(backend2, graph);
if (copy.buffer == NULL) {
return false;
@ -1837,28 +1838,45 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
assert(g1->n_nodes == g2->n_nodes);
for (int i = 0; i < g1->n_nodes; i++) {
struct ggml_tensor * t1 = g1->nodes[i];
struct ggml_tensor * t2 = g2->nodes[i];
if (test_node != nullptr) {
// Compute the whole graph and only test the output for a specific tensor
ggml_backend_graph_compute(backend1, g1);
ggml_backend_graph_compute(backend2, g2);
assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
ggml_backend_graph_compute(backend1, &g1v);
ggml_backend_graph_compute(backend2, &g2v);
if (ggml_is_view_op(t1->op)) {
continue;
int test_node_idx = -1;
for (int i = 0; i < g1->n_nodes; i++) {
struct ggml_tensor * t1 = g1->nodes[i];
if (t1 == test_node) {
test_node_idx = i;
break;
}
}
GGML_ASSERT(test_node_idx != -1);
// compare results, calculate rms etc
if (!callback(i, t1, t2, user_data)) {
break;
callback(test_node_idx, g1->nodes[test_node_idx], g2->nodes[test_node_idx], user_data);
} else {
for (int i = 0; i < g1->n_nodes; i++) {
struct ggml_tensor * t1 = g1->nodes[i];
struct ggml_tensor * t2 = g2->nodes[i];
assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
ggml_backend_graph_compute(backend1, &g1v);
ggml_backend_graph_compute(backend2, &g2v);
if (ggml_is_view_op(t1->op)) {
continue;
}
// compare results, calculate rms etc
if (!callback(i, t1, t2, user_data)) {
break;
}
}
}
ggml_backend_graph_copy_free(copy);
return true;

73
ggml/src/ggml-cann/aclnn_ops.cpp
View File

@ -65,8 +65,9 @@
#include <aclnnop/aclnn_eq_tensor.h>
#include <aclnnop/aclnn_gt_scalar.h>
#include <aclnnop/aclnn_pow.h>
#include <aclnnop/aclnn_grouped_matmul_v2.h>
#include <aclnnop/aclnn_grouped_matmul_v3.h>
#include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
#include <aclnnop/aclnn_zero.h>
#include <float.h>
#include <cmath>
@ -804,10 +805,11 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer,
nb[i] = nb[i - 1] * ne[i - 1];
}
ggml_cann_async_memset(ctx, buffer, n_bytes, 0);
aclTensor* zero =
ggml_cann_create_tensor(buffer, type, type_size, ne, nb, dims);
GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, zero);
return zero;
GGML_UNUSED(n_bytes);
}
/**
@ -2654,6 +2656,67 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
memcpy(ori_src0_nb, cast_nb, sizeof(ori_src0_nb));
}
#ifdef ASCEND_310P
ggml_tensor src0_row = *src0;
ggml_tensor src1_row = *src1;
ggml_tensor dst_row = *dst;
if (src0->type == GGML_TYPE_F16) {
src0_row.type = GGML_TYPE_F32;
}
// src0_row [D, M, 1, 1] weight without permute
src0_row.ne[2] = 1;
src0_row.ne[3] = 1;
src0_row.nb[0] = ori_src0_nb[0];
src0_row.nb[1] = ori_src0_nb[1];
src0_row.nb[2] = ori_src0_nb[1];
src0_row.nb[3] = ori_src0_nb[1];
// src1_row [D, 1, 1, 1] -> input
src1_row.ne[1] = 1;
src1_row.ne[2] = 1;
src1_row.ne[3] = 1;
src1_row.nb[2] = nb11;
src1_row.nb[3] = nb11;
// dst_row [M, 1, 1, 1] -> out
dst_row.ne[1] = 1;
dst_row.ne[2] = 1;
dst_row.ne[3] = 1;
dst_row.nb[2] = nb1;
dst_row.nb[3] = nb1;
//create weight for one row
for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
for (int64_t id = 0; id < n_ids; id++) {
// expert index
int32_t i02 = *(int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
GGML_ASSERT(i02 >= 0 && i02 < n_as);
// If B = 1 (broadcast), always use 0; otherwise, use id.
int64_t i11 = (ne11 == 1 ? 0 : id);
int64_t i12 = iid1;
int64_t i1 = id;
int64_t i2 = i12;
void* src0_tmp_ptr = src0_original + i02*ori_src0_nb[2];
void* src1_tmp_ptr = src1_original + i11*nb11 + i12*nb12;
void* dst_tmp_ptr = dst_original + i1*nb1 + i2*nb2;
src0_row.data = src0_tmp_ptr;
src1_row.data = src1_tmp_ptr;
dst_row.data = dst_tmp_ptr;
dst_row.src[0] = &src0_row;
dst_row.src[1] = &src1_row;
ggml_cann_mul_mat(ctx, &dst_row);
}
}
return;
#endif
std::vector<aclTensor*> src0_tensor_vec;
std::vector<aclTensor*> src1_tensor_vec;
std::vector<aclTensor*> dst_tensor_vec;
@ -2701,9 +2764,9 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
}
size_t GROUP_SIZE = 128;
// GroupedMatmulV2 required tensor_list.size < 128
// GroupedMatmulV3 required tensor_list.size < 128
for (size_t i = 0; i < src0_tensor_vec.size(); i += GROUP_SIZE) {
// split and call GroupedMatmulV2
// split and call GroupedMatmulV3
size_t end = std::min(i + GROUP_SIZE, src0_tensor_vec.size());
std::vector<aclTensor*> src0_tensor_vec_split(src0_tensor_vec.begin() + i, src0_tensor_vec.begin() + end);
std::vector<aclTensor*> src1_tensor_vec_split(src1_tensor_vec.begin() + i, src1_tensor_vec.begin() + end);
@ -2713,7 +2776,7 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
aclTensorList* src1_tensor_list = aclCreateTensorList(src1_tensor_vec_split.data(), src1_tensor_vec_split.size());
aclTensorList* dst_tensor_list = aclCreateTensorList(dst_tensor_vec_split.data(), dst_tensor_vec_split.size());
GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV2, src1_tensor_list, src0_tensor_list,
GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV3, src1_tensor_list, src0_tensor_list,
nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, -1, dst_tensor_list);
ggml_cann_release_resources(ctx, src0_tensor_list, src1_tensor_list, dst_tensor_list);

2
ggml/src/ggml-cann/common.h
View File

@ -359,7 +359,7 @@ struct ggml_backend_cann_context {
ggml_cann_set_device(device);
description = aclrtGetSocName();
bool async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
device, async_mode ? "ON" : "OFF");
}

13
ggml/src/ggml-cann/ggml-cann.cpp
View File

@ -2086,6 +2086,12 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
return false;
}
} break;
case GGML_OP_SET_ROWS:
{
// TODO: add support
// ref: https://github.com/ggml-org/llama.cpp/pull/14274
return false;
} break;
case GGML_OP_CPY: {
ggml_tensor *src = op->src[0];
if ((op->type != GGML_TYPE_F32 && op->type != GGML_TYPE_F16) ||
@ -2187,7 +2193,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
case GGML_OP_SQRT:
case GGML_OP_CLAMP:
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_SOFT_MAX:
case GGML_OP_SUM_ROWS:
case GGML_OP_ARGSORT:
case GGML_OP_ACC:
@ -2205,6 +2210,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
case GGML_OP_PAD_REFLECT_1D:
case GGML_OP_COUNT_EQUAL:
return true;
case GGML_OP_SOFT_MAX:
// TODO: support broadcast
// ref: https://github.com/ggml-org/llama.cpp/pull/14435
return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
case GGML_OP_FLASH_ATTN_EXT:{
// derived from [ggml-cuda.cu]
if(op->src[1]->type != GGML_TYPE_F16 || op->src[2]->type != GGML_TYPE_F16){
@ -2227,6 +2236,8 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
// DeepSeek MLA
return false;
}
// TODO: support broadcast
// ref: https://github.com/ggml-org/llama.cpp/pull/14435
if (op->src[0]->ne[3] != 1) {
return false;
}

7
ggml/src/ggml-cpu/CMakeLists.txt
View File

@ -5,7 +5,7 @@ function(ggml_add_cpu_backend_features cpu_name arch)
# build, using set_source_files_properties() to set the arch flags is not possible
set(GGML_CPU_FEATS_NAME ${cpu_name}-feats)
add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/arch/${arch}/cpu-feats.cpp)
target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . ../include)
target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARGN})
target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
@ -589,4 +589,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
if (EMSCRIPTEN)
set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")
endif()
if (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
# The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"
target_compile_options(${GGML_CPU_NAME} PRIVATE "-fno-associative-math")
endif()
endfunction()

39
ggml/src/ggml-cpu/ggml-cpu.c
View File

@ -195,6 +195,7 @@ typedef pthread_t ggml_thread_t;
static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
[GGML_TYPE_F32] = {
.from_float = (ggml_from_float_t) ggml_cpu_fp32_to_fp32,
.vec_dot = (ggml_vec_dot_t) ggml_vec_dot_f32,
.vec_dot_type = GGML_TYPE_F32,
.nrows = 1,
@ -1192,7 +1193,7 @@ static void ggml_compute_forward_mul_mat_one_chunk(
}
}
static void ggml_compute_forward_mul_mat(
void ggml_compute_forward_mul_mat(
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
@ -1817,6 +1818,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
{
ggml_compute_forward_get_rows_back(params, tensor);
} break;
case GGML_OP_SET_ROWS:
{
ggml_compute_forward_set_rows(params, tensor);
} break;
case GGML_OP_DIAG:
{
ggml_compute_forward_diag(params, tensor);
@ -1861,6 +1866,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
{
ggml_compute_forward_im2col_back_f32(params, tensor);
} break;
case GGML_OP_CONV_2D:
{
ggml_compute_forward_conv_2d(params, tensor);
} break;
case GGML_OP_CONV_2D_DW:
{
ggml_compute_forward_conv_2d_dw(params, tensor);
@ -1944,6 +1953,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
{
ggml_compute_forward_unary(params, tensor);
} break;
case GGML_OP_GLU:
{
ggml_compute_forward_glu(params, tensor);
} break;
case GGML_OP_GET_REL_POS:
{
ggml_compute_forward_get_rel_pos(params, tensor);
@ -2154,6 +2167,20 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
GGML_ABORT("fatal error");
}
break;
case GGML_OP_GLU:
switch (ggml_get_glu_op(node)) {
case GGML_GLU_OP_REGLU:
case GGML_GLU_OP_GEGLU:
case GGML_GLU_OP_SWIGLU:
case GGML_GLU_OP_GEGLU_ERF:
case GGML_GLU_OP_GEGLU_QUICK:
{
n_tasks = n_threads;
} break;
default:
GGML_ABORT("fatal error");
}
break;
case GGML_OP_SILU_BACK:
case GGML_OP_MUL:
case GGML_OP_DIV:
@ -2170,6 +2197,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
n_tasks = n_threads;
} break;
case GGML_OP_GET_ROWS:
case GGML_OP_SET_ROWS:
{
// FIXME: get_rows can use additional threads, but the cost of launching additional threads
// decreases performance with GPU offloading
@ -2206,6 +2234,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
} break;
case GGML_OP_IM2COL:
case GGML_OP_IM2COL_BACK:
case GGML_OP_CONV_2D:
case GGML_OP_CONV_2D_DW:
case GGML_OP_CONV_TRANSPOSE_1D:
case GGML_OP_CONV_TRANSPOSE_2D:
@ -2724,6 +2753,10 @@ struct ggml_cplan ggml_graph_plan(
GGML_ABORT("fatal error");
}
} break;
case GGML_OP_CONV_2D:
{
cur = GGML_IM2COL_WORK_SIZE;
} break;
case GGML_OP_CONV_TRANSPOSE_2D:
{
const int64_t ne00 = node->src[0]->ne[0]; // W
@ -3124,6 +3157,10 @@ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct g
return ggml_graph_compute(cgraph, &cplan);
}
void ggml_cpu_fp32_to_fp32(const float * x, float * y, int64_t n) {
memcpy(y, x, n * sizeof(float));
}
void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
int64_t i = 0;
#if defined(__F16C__)

1
ggml/src/ggml-cpu/ggml-cpu.cpp
View File

@ -416,6 +416,7 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
switch (op->op) {
case GGML_OP_CPY:
case GGML_OP_SET_ROWS:
return
op->type != GGML_TYPE_IQ3_XXS &&
op->type != GGML_TYPE_IQ3_S &&

1426
ggml/src/ggml-cpu/ops.cpp
View File

File diff suppressed because it is too large Load Diff

7
ggml/src/ggml-cpu/ops.h
View File

@ -20,6 +20,9 @@
static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
// Work buffer size for im2col operations in CONV2D
#define GGML_IM2COL_WORK_SIZE (16 * 1024 * 1024)
#ifdef __cplusplus
extern "C" {
#endif
@ -53,6 +56,7 @@ void ggml_compute_forward_permute(const struct ggml_compute_params * params, str
void ggml_compute_forward_transpose(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_get_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_get_rows_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_set_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_diag(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_diag_mask_inf(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_diag_mask_zero(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@ -64,6 +68,7 @@ void ggml_compute_forward_clamp(const struct ggml_compute_params * params, struc
void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_2d_dw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@ -93,6 +98,7 @@ void ggml_compute_forward_ssm_scan(const struct ggml_compute_params * params, st
void ggml_compute_forward_win_part(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_win_unpart(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_unary(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_glu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_get_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rwkv_wkv6(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@ -105,6 +111,7 @@ void ggml_compute_forward_custom(const struct ggml_compute_params * params, stru
void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
#ifdef __cplusplus
}

2
ggml/src/ggml-cpu/simd-mappings.h
View File

@ -189,7 +189,7 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
#define GGML_F32xt_LOAD(...) GGML_F32xt_LOAD_IMPL(DEFAULT_PG, __VA_ARGS__)
#define GGML_F32xt_STORE_IMPL(pg,a,b) svst1_f32(pg, a, b)
#define GGML_F32xt_STORE(...) GGML_F32xt_STORE_IMPL(DEFAULT_PG, __VA_ARGS__)
#define GGML_F32xt_FMA_IMPL(pg, a, b, c) svmad_f32_m(pg, a, b, c)
#define GGML_F32xt_FMA_IMPL(pg, a, b, c) svmad_f32_m(pg, b, c, a)
#define GGML_F32xt_FMA(...) GGML_F32xt_FMA_IMPL(DEFAULT_PG, __VA_ARGS__)
#define GGML_F32xt_ADD_IMPL(pg, a, b) svadd_f32_m(pg, a, b)
#define GGML_F32xt_ADD(...) GGML_F32xt_ADD_IMPL(DEFAULT_PG, __VA_ARGS__)

42
ggml/src/ggml-cpu/vec.cpp
View File

@ -37,35 +37,35 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
for (int i = 0; i < np; i += ggml_f32_step) {
ax1 = GGML_F32_VEC_LOAD(x + i);
ay1 = GGML_F32_VEC_LOAD(y + i);
sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1);
sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
sum2 = GGML_F32_VEC_FMA(ax2, ay2, sum2);
sum2 = GGML_F32_VEC_FMA(sum2, ax2, ay2);
ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
sum3 = GGML_F32_VEC_FMA(ax3, ay3, sum3);
sum3 = GGML_F32_VEC_FMA(sum3, ax3, ay3);
ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
sum4 = GGML_F32_VEC_FMA(ax4, ay4, sum4);
sum4 = GGML_F32_VEC_FMA(sum4, ax4, ay4);
ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
sum5 = GGML_F32_VEC_FMA(ax5, ay5, sum5);
sum5 = GGML_F32_VEC_FMA(sum5, ax5, ay5);
ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
sum6 = GGML_F32_VEC_FMA(ax6, ay6, sum6);
sum6 = GGML_F32_VEC_FMA(sum6, ax6, ay6);
ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
sum7 = GGML_F32_VEC_FMA(ax7, ay7, sum7);
sum7 = GGML_F32_VEC_FMA(sum7, ax7, ay7);
ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
sum8 = GGML_F32_VEC_FMA(ax8, ay8, sum8);
sum8 = GGML_F32_VEC_FMA(sum8, ax8, ay8);
}
// leftovers
// Since 8 unrolls are done in above loop, leftovers lie in range [0, ggml_f32_step] which is handled in below loop
@ -73,7 +73,7 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
for (int i = np; i < np2; i += ggml_f32_epr) {
ax1 = GGML_F32_VEC_LOAD(x + i);
ay1 = GGML_F32_VEC_LOAD(y + i);
sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1);
sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
}
// maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
if (np2 < n) {
@ -254,6 +254,30 @@ void ggml_vec_silu_f32(const int n, float * y, const float * x) {
}
}
void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float * g) {
int i = 0;
#if defined(__AVX512F__) && defined(__AVX512DQ__)
for (; i + 15 < n; i += 16) {
_mm512_storeu_ps(y + i, _mm512_mul_ps(ggml_v_silu(_mm512_loadu_ps(x + i)), _mm512_loadu_ps(g + i)));
}
#elif defined(__AVX2__) && defined(__FMA__)
for (; i + 7 < n; i += 8) {
_mm256_storeu_ps(y + i, _mm256_mul_ps(ggml_v_silu(_mm256_loadu_ps(x + i)), _mm256_loadu_ps(g + i)));
}
#elif defined(__SSE2__)
for (; i + 3 < n; i += 4) {
_mm_storeu_ps(y + i, _mm_mul_ps(ggml_v_silu(_mm_loadu_ps(x + i)), _mm_loadu_ps(g + i)));
}
#elif defined(__ARM_NEON) && defined(__aarch64__)
for (; i + 3 < n; i += 4) {
vst1q_f32(y + i, vmulq_f32(ggml_v_silu(vld1q_f32(x + i)), vld1q_f32(g + i)));
}
#endif
for (; i < n; ++i) {
y[i] = ggml_silu_f32(x[i]) * g[i];
}
}
ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max) {
int i = 0;
ggml_float sum = 0;

112
ggml/src/ggml-cpu/vec.h
View File

@ -163,49 +163,49 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
ax1 = GGML_F32_VEC_LOAD(x + i);
ay1 = GGML_F32_VEC_LOAD(y + i);
ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1);
ay1 = GGML_F32_VEC_FMA(ay1, ax1, vx);
GGML_F32_VEC_STORE(y + i, ay1);
ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
ay2 = GGML_F32_VEC_FMA(ax2, vx, ay2);
ay2 = GGML_F32_VEC_FMA(ay2, ax2, vx);
GGML_F32_VEC_STORE(y + i + 1*ggml_f32_epr, ay2);
ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
ay3 = GGML_F32_VEC_FMA(ax3, vx, ay3);
ay3 = GGML_F32_VEC_FMA(ay3, ax3, vx);
GGML_F32_VEC_STORE(y + i + 2*ggml_f32_epr, ay3);
ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
ay4 = GGML_F32_VEC_FMA(ax4, vx, ay4);
ay4 = GGML_F32_VEC_FMA(ay4, ax4, vx);
GGML_F32_VEC_STORE(y + i + 3*ggml_f32_epr, ay4);
ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
ay5 = GGML_F32_VEC_FMA(ax5, vx, ay5);
ay5 = GGML_F32_VEC_FMA(ay5, ax5, vx);
GGML_F32_VEC_STORE(y + i + 4*ggml_f32_epr, ay5);
ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
ay6 = GGML_F32_VEC_FMA(ax6, vx, ay6);
ay6 = GGML_F32_VEC_FMA(ay6, ax6, vx);
GGML_F32_VEC_STORE(y + i + 5*ggml_f32_epr, ay6);
ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
ay7 = GGML_F32_VEC_FMA(ax7, vx, ay7);
ay7 = GGML_F32_VEC_FMA(ay7, ax7, vx);
GGML_F32_VEC_STORE(y + i + 6*ggml_f32_epr, ay7);
ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
ay8 = GGML_F32_VEC_FMA(ax8, vx, ay8);
ay8 = GGML_F32_VEC_FMA(ay8, ax8, vx);
GGML_F32_VEC_STORE(y + i + 7*ggml_f32_epr, ay8);
}
@ -215,7 +215,7 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
for (int i = np; i < np2; i += ggml_f32_epr) {
ax1 = GGML_F32_VEC_LOAD(x + i);
ay1 = GGML_F32_VEC_LOAD(y + i);
ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1);
ay1 = GGML_F32_VEC_FMA(ay1, ax1, vx);
GGML_F32_VEC_STORE(y + i, ay1);
}
@ -905,6 +905,100 @@ inline static void ggml_vec_silu_backward_f16(const int n, ggml_fp16_t * dx, con
}
}
inline static void ggml_vec_reglu_f32 (const int n, float * y, const float * x, const float * g) {
for (int i = 0; i < n; ++i) {
y[i] = (x[i] > 0.f) ? x[i] * g[i] : 0.f;
}
}
inline static void ggml_vec_reglu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
for (int i = 0; i < n; ++i) {
float v = GGML_CPU_FP16_TO_FP32(x[i]);
y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? v * GGML_CPU_FP16_TO_FP32(g[i]) : 0.f);
}
}
#ifdef GGML_GELU_FP16
inline static void ggml_vec_geglu_f32(const int n, float * y, const float * x, const float * g) {
uint16_t t;
for (int i = 0; i < n; ++i) {
if (x[i] <= -10.0f) {
y[i] = 0.0f;
} else if (x[i] >= 10.0f) {
y[i] = x[i] * g[i];
} else {
ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
memcpy(&t, &fp16, sizeof(uint16_t));
y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_f16[t]) * g[i];
}
}
}
#else
inline static void ggml_vec_geglu_f32(const int n, float * y, const float * x, const float * g) {
for (int i = 0; i < n; ++i) {
y[i] = ggml_gelu_f32(x[i]) * g[i];
}
}
#endif
inline static void ggml_vec_geglu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
const uint16_t * i16 = (const uint16_t *) x;
for (int i = 0; i < n; ++i) {
float v = GGML_CPU_FP16_TO_FP32(g[i]);
y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(ggml_table_gelu_f16[i16[i]]) * v);
}
}
void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float * g);
inline static void ggml_vec_swiglu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
for (int i = 0; i < n; ++i) {
float v = GGML_CPU_FP16_TO_FP32(x[i]);
float w = GGML_CPU_FP16_TO_FP32(g[i]);
y[i] = GGML_CPU_FP32_TO_FP16((v/(1.0f + expf(-v))) * w);
}
}
inline static void ggml_vec_geglu_erf_f32(const int n, float * y, const float * x, const float * g) {
for (int i = 0; i < n; ++i) {
float xi = x[i];
y[i] = 0.5f * xi * (1.0f + erff(xi*SQRT_2_INV)) * g[i];
}
}
inline static void ggml_vec_geglu_erf_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
for (int i = 0; i < n; ++i) {
float xi = GGML_CPU_FP16_TO_FP32(x[i]);
float gi = GGML_CPU_FP16_TO_FP32(g[i]);
y[i] = GGML_CPU_FP32_TO_FP16(0.5f * xi * (1.0f + erff(xi*SQRT_2_INV)) * gi);
}
}
#ifdef GGML_GELU_QUICK_FP16
inline static void ggml_vec_geglu_quick_f32(const int n, float * y, const float * x, const float * g) {
uint16_t t;
for (int i = 0; i < n; ++i) {
ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
memcpy(&t, &fp16, sizeof(uint16_t));
y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]) * g[i];
}
}
#else
inline static void ggml_vec_geglu_quick_f32(const int n, float * y, const float * x, const float * g) {
for (int i = 0; i < n; ++i) {
y[i] = ggml_gelu_quick_f32(x[i]) * g[i];
}
}
#endif
inline static void ggml_vec_geglu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
const uint16_t * i16 = (const uint16_t *) x;
for (int i = 0; i < n; ++i) {
float v = GGML_CPU_FP16_TO_FP32(g[i]);
y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[i16[i]]) * v);
}
}
inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) {
#ifndef GGML_USE_ACCELERATE
ggml_float sum = 0.0;

14
ggml/src/ggml-cuda/common.cuh
View File

@ -175,6 +175,20 @@ static const char * cu_get_error_str(CUresult err) {
#define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
#endif
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
#define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes) \
do { \
static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false}; \
const int id = ggml_cuda_get_device(); \
if (!shared_memory_limit_raised[id]) { \
CUDA_CHECK(cudaFuncSetAttribute(kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes)); \
shared_memory_limit_raised[id] = true; \
} \
} while (0)
#else
#define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes) do {} while (0)
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
#if CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA)
#define GGML_CUDA_ASSUME(x) __builtin_assume(x)
#else

22
ggml/src/ggml-cuda/convert.cu
View File

@ -728,3 +728,25 @@ to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
return nullptr;
}
}
to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type) {
switch (type) {
case GGML_TYPE_F32:
return convert_unary_cuda<float, nv_bfloat16>;
case GGML_TYPE_F16:
return convert_unary_cuda<half, nv_bfloat16>;
default:
return nullptr;
}
}
to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type) {
switch (type) {
case GGML_TYPE_F16:
return convert_unary_cuda<half, float>;
case GGML_TYPE_BF16:
return convert_unary_cuda<nv_bfloat16, float>;
default:
return nullptr;
}
}

5
ggml/src/ggml-cuda/convert.cuh
View File

@ -22,5 +22,10 @@ using to_t_nc_cuda_t = void (*)(const void * x, T * y,
int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03,
int64_t s01, int64_t s02, int64_t s03, cudaStream_t stream);
typedef to_t_nc_cuda_t<float> to_fp32_nc_cuda_t;
typedef to_t_nc_cuda_t<half> to_fp16_nc_cuda_t;
typedef to_t_nc_cuda_t<nv_bfloat16> to_bf16_nc_cuda_t;
to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type);
to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type);
to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type);

16
ggml/src/ggml-cuda/cross-entropy-loss.cu
View File

@ -123,13 +123,7 @@ void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor *
ggml_cuda_pool_alloc<float> dst_tmp(pool, blocks_num.x);
if (nbytes_shared <= smpbo) {
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
if (!shared_memory_limit_raised[id]) {
CUDA_CHECK(cudaFuncSetAttribute(cross_entropy_loss_f32<true>, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo));
shared_memory_limit_raised[id] = true;
}
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
CUDA_SET_SHARED_MEMORY_LIMIT((cross_entropy_loss_f32<true>), smpbo);
cross_entropy_loss_f32<true><<<blocks_num, blocks_dim, nbytes_shared, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
} else {
cross_entropy_loss_f32<false><<<blocks_num, blocks_dim, 0, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
@ -175,13 +169,7 @@ void ggml_cuda_cross_entropy_loss_back(ggml_backend_cuda_context & ctx, ggml_ten
const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
if (nbytes_shared <= smpbo) {
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
if (!shared_memory_limit_raised[id]) {
CUDA_CHECK(cudaFuncSetAttribute(cross_entropy_loss_back_f32<true>, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo));
shared_memory_limit_raised[id] = true;
}
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
CUDA_SET_SHARED_MEMORY_LIMIT((cross_entropy_loss_back_f32<true>), smpbo);
cross_entropy_loss_back_f32<true><<<blocks_num, blocks_dim, nbytes_shared, stream>>>(grad_d, src0f_d, src1f_d, dst_d, ne00);
} else {
cross_entropy_loss_back_f32<false><<<blocks_num, blocks_dim, 0, stream>>>(grad_d, src0f_d, src1f_d, dst_d, ne00);

5
ggml/src/ggml-cuda/fattn-common.cuh
View File

@ -32,7 +32,9 @@ typedef void (* fattn_kernel_t)(
const int ne12,
const int ne13,
const int ne31,
const int ne32,
const int nb31,
const int nb32,
const int nb01,
const int nb02,
const int nb03,
@ -851,7 +853,8 @@ void launch_fattn(
scale, max_bias, m0, m1, n_head_log2, logit_softcap,
Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
K->ne[0], K->ne[1], K->ne[2], K->ne[3],
mask ? mask->ne[1] : 0, mask ? mask->nb[1] : 0,
mask ? mask->ne[1] : 0, mask ? mask->ne[2] : 0,
mask ? mask->nb[1] : 0, mask ? mask->nb[2] : 0,
Q->nb[1], Q->nb[2], Q->nb[3],
nb11, nb12, nb13,
nb21, nb22, nb23,

12
ggml/src/ggml-cuda/fattn-mma-f16.cuh
View File

@ -1223,7 +1223,9 @@ static __global__ void flash_attn_ext_f16(
const int ne12,
const int ne13,
const int ne31,
const int ne32,
const int nb31,
const int nb32,
const int nb01,
const int nb02,
const int nb03,
@ -1288,7 +1290,8 @@ static __global__ void flash_attn_ext_f16(
const float2 * Q_f2 = (const float2 *) (Q + nb02* channel*ncols2);
const half2 * K_h2 = (const half2 *) (K + nb12*(channel*ncols2 / gqa_ratio));
const half2 * mask_h2 = ncols2 > 1 || mask ? (const half2 *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
const half2 * mask_h2 = ncols2 == 1 && !mask ? nullptr :
(const half2 *) (mask + nb32*(channel % ne32) + nb31*jt*ncols1);
float2 * dstk = ((float2 *) dst) + channel*(ncols2 * DV/2);
const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
@ -1327,7 +1330,8 @@ static __global__ void flash_attn_ext_f16(
const float2 * Q_f2 = (const float2 *) (Q + nb02* channel*ncols2);
const half2 * K_h2 = (const half2 *) (K + nb12*(channel*ncols2 / gqa_ratio));
const half2 * mask_h2 = ncols2 > 1 || mask ? (const half2 *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
const half2 * mask_h2 = ncols2 == 1 && !mask ? nullptr :
(const half2 *) (mask + nb32*(channel % ne32) + nb31*jt*ncols1);
float2 * dstk = ((float2 *) dst) + channel*(ncols2 * DV/2);
const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
@ -1348,8 +1352,8 @@ static __global__ void flash_attn_ext_f16(
GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne00);
GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); GGML_UNUSED(ne10);
GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); GGML_UNUSED(nb21);
GGML_UNUSED(nb22); GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
GGML_UNUSED(ne2); GGML_UNUSED(ne3);

10
ggml/src/ggml-cuda/fattn-tile-f16.cu
View File

@ -6,7 +6,7 @@
template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(nwarps*WARP_SIZE, 1)
__launch_bounds__(nwarps*WARP_SIZE, 2)
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_tile_ext_f16(
const char * __restrict__ Q,
@ -30,7 +30,9 @@ static __global__ void flash_attn_tile_ext_f16(
const int ne12,
const int ne13,
const int ne31,
const int ne32,
const int nb31,
const int nb32,
const int nb01,
const int nb02,
const int nb03,
@ -64,7 +66,7 @@ static __global__ void flash_attn_tile_ext_f16(
const float2 * Q_f2 = (const float2 *) (Q + nb02* blockIdx.z + nb01*ic0);
const half2 * K_h2 = (const half2 *) (K + nb12*(blockIdx.z / gqa_ratio));
const half2 * V_h2 = (const half2 *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
const half * maskh = (const half *) mask + ne11*ic0;
const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
const int stride_KV2 = nb11 / sizeof(half2);
@ -288,8 +290,8 @@ static __global__ void flash_attn_tile_ext_f16(
GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);

10
ggml/src/ggml-cuda/fattn-tile-f32.cu
View File

@ -6,7 +6,7 @@
template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(nwarps*WARP_SIZE, 1)
__launch_bounds__(nwarps*WARP_SIZE, 2)
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_tile_ext_f32(
const char * __restrict__ Q,
@ -30,7 +30,9 @@ static __global__ void flash_attn_tile_ext_f32(
const int ne12,
const int ne13,
const int ne31,
const int ne32,
const int nb31,
const int nb32,
const int nb01,
const int nb02,
const int nb03,
@ -58,8 +60,8 @@ static __global__ void flash_attn_tile_ext_f32(
GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
@ -76,7 +78,7 @@ static __global__ void flash_attn_tile_ext_f32(
const float2 * Q_f2 = (const float2 *) (Q + nb02* blockIdx.z + nb01*ic0);
const half2 * K_h2 = (const half2 *) (K + nb12*(blockIdx.z / gqa_ratio));
const half2 * V_h2 = (const half2 *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
const half * maskh = (const half *) mask + ne11*ic0;
const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
const int stride_KV2 = nb11 / sizeof(half2);

8
ggml/src/ggml-cuda/fattn-vec-f16.cuh
View File

@ -27,7 +27,9 @@ static __global__ void flash_attn_vec_ext_f16(
const int ne12,
const int ne13,
const int ne31,
const int ne32,
const int nb31,
const int nb32,
const int nb01,
const int nb02,
const int nb03,
@ -68,7 +70,7 @@ static __global__ void flash_attn_vec_ext_f16(
K += nb12*(blockIdx.z / gqa_ratio);
V += nb22*(blockIdx.z / gqa_ratio);
const half * maskh = (const half *) mask + ne11*ic0;
const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
const half slopeh = __float2half(slopef);
@ -342,8 +344,8 @@ static __global__ void flash_attn_vec_ext_f16(
GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);

9
ggml/src/ggml-cuda/fattn-vec-f32.cuh
View File

@ -27,7 +27,9 @@ static __global__ void flash_attn_vec_ext_f32(
const int ne12,
const int ne13,
const int ne31,
const int ne32,
const int nb31,
const int nb32,
const int nb01,
const int nb02,
const int nb03,
@ -51,8 +53,8 @@ static __global__ void flash_attn_vec_ext_f32(
GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
@ -79,7 +81,8 @@ static __global__ void flash_attn_vec_ext_f32(
Q += nb02* blockIdx.z + nb01*ic0;
K += nb12*(blockIdx.z / gqa_ratio);
V += nb22*(blockIdx.z / gqa_ratio); // K and V have same shape
const half * maskh = (const half *) mask + ne11*ic0;
const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);

14
ggml/src/ggml-cuda/fattn-wmma-f16.cu
View File

@ -46,7 +46,9 @@ static __global__ void flash_attn_ext_f16(
const int ne12,
const int ne13,
const int ne31,
const int ne32,
const int nb31,
const int nb32,
const int nb01,
const int nb02,
const int nb03,
@ -94,11 +96,11 @@ static __global__ void flash_attn_ext_f16(
constexpr int kqar = sizeof(KQ_acc_t)/sizeof(half);
const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
const float * Q_f = (const float *) (Q + nb02* blockIdx.z + nb01*ic0);
const half * K_h = (const half *) (K + nb12*(blockIdx.z / gqa_ratio));
const half * V_h = (const half *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
const half * maskh = (const half *) mask + (nb31/sizeof(half))* ic0;
const half2 * mask2 = (const half2 *) mask + (nb31/sizeof(half))*(ic0/2);
const float * Q_f = (const float *) (Q + nb02* blockIdx.z + nb01*ic0);
const half * K_h = (const half *) (K + nb12*(blockIdx.z / gqa_ratio));
const half * V_h = (const half *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
const half2 * mask2 = (const half2 *) maskh;
const int stride_Q = nb01 / sizeof(float);
const int stride_KV = nb11 / sizeof(half);
@ -440,7 +442,7 @@ static __global__ void flash_attn_ext_f16(
GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
GGML_UNUSED(ne31); GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3);

8
ggml/src/ggml-cuda/getrows.cu
View File

@ -168,6 +168,10 @@ static void ggml_cuda_get_rows_switch_src0_type(
get_rows_cuda_float((const float *) src0_d, src1_d, dst_d,
ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
break;
case GGML_TYPE_I32:
get_rows_cuda_float((const int32_t *) src0_d, src1_d, dst_d,
ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
break;
case GGML_TYPE_BF16:
get_rows_cuda_float((const nv_bfloat16 *) src0_d, src1_d, dst_d,
ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
@ -210,6 +214,10 @@ void get_rows_cuda(
ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (float *) dst_d,
ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
break;
case GGML_TYPE_I32:
ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (int32_t *) dst_d,
ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
break;
case GGML_TYPE_F16:
ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (half *) dst_d,
ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);

263
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@ -1749,7 +1749,7 @@ static void ggml_cuda_op_mul_mat(
}
static __global__ void k_compute_batched_ptrs(
const half * src0_as_f16, const half * src1_as_f16, char * dst,
const void * src0_as_f16, const void * src1_as_f16, char * dst,
const void ** ptrs_src, void ** ptrs_dst,
int64_t ne12, int64_t ne13,
int64_t ne23,
@ -1772,83 +1772,131 @@ static __global__ void k_compute_batched_ptrs(
ptrs_dst[0*ne23 + i12 + i13*ne12] = ( char *) dst + i12*nbd2 + i13*nbd3;
}
static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
// Type traits for mapping ggml types to CUDA/cuBLAS types
template<ggml_type T>
struct batched_mul_mat_traits;
template<>
struct batched_mul_mat_traits<GGML_TYPE_F32> {
using cuda_type = float;
static inline const cublasComputeType_t compute_type = CUBLAS_COMPUTE_32F;
static inline const cudaDataType_t data_type = CUDA_R_32F;
static inline const ggml_type ggml_type_val = GGML_TYPE_F32;
static inline const float alpha = 1.0f;
static inline const float beta = 0.0f;
static inline const void* get_alpha() { static const float val = alpha; return &val; }
static inline const void* get_beta() { static const float val = beta; return &val; }
static inline auto get_nc_converter(ggml_type src_type) { return ggml_get_to_fp32_nc_cuda(src_type); }
};
template<>
struct batched_mul_mat_traits<GGML_TYPE_BF16> {
using cuda_type = nv_bfloat16;
static inline const cublasComputeType_t compute_type = CUBLAS_COMPUTE_32F;
static inline const cudaDataType_t data_type = CUDA_R_16BF;
static inline const ggml_type ggml_type_val = GGML_TYPE_BF16;
static inline const float alpha = 1.0f;
static inline const float beta = 0.0f;
static inline const void* get_alpha() { static const float val = alpha; return &val; }
static inline const void* get_beta() { static const float val = beta; return &val; }
static inline auto get_nc_converter(ggml_type src_type) { return ggml_get_to_bf16_nc_cuda(src_type); }
};
template<>
struct batched_mul_mat_traits<GGML_TYPE_F16> {
using cuda_type = half;
static inline const cublasComputeType_t compute_type = CUBLAS_COMPUTE_16F;
static inline const cudaDataType_t data_type = CUDA_R_16F;
static inline const ggml_type ggml_type_val = GGML_TYPE_F16;
static inline const half alpha = 1.0;
static inline const half beta = 0.0;
static inline const void* get_alpha() { static const half val = alpha; return &val; }
static inline const void* get_beta() { static const half val = beta; return &val; }
static inline auto get_nc_converter(ggml_type src_type) { return ggml_get_to_fp16_nc_cuda(src_type); }
};
template<ggml_type src0_type>
static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
using traits = batched_mul_mat_traits<src0_type>;
using cuda_t = typename traits::cuda_type;
GGML_ASSERT(!ggml_is_transposed(src0));
GGML_ASSERT(!ggml_is_transposed(src1));
GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft));
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src0->type == src0_type);
GGML_ASSERT(ggml_is_contiguous(dst));
// Byte offsets and tensor dimensions are currently used in an inconsistent way for dst.
// As long as dst is contiguous this does not matter though.
GGML_ASSERT(ggml_is_contiguous(dst));
GGML_TENSOR_BINARY_OP_LOCALS
const int64_t ne_dst = ggml_nelements(dst);
cudaStream_t main_stream = ctx.stream();
CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));
const half * src0_f16 = (const half *) src0->data;
float * dst_ddf = (float *) dst->data;
const half * src1_f16 = (const half *) src1->data;
const size_t ts_src1 = ggml_type_size(src1->type);
GGML_ASSERT(nb10 == ts_src1);
int64_t s11 = nb11 / ts_src1;
int64_t s12 = nb12 / ts_src1;
int64_t s13 = nb13 / ts_src1;
ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());
// convert src1 to fp16
if (src1->type != GGML_TYPE_F16) {
const to_fp16_nc_cuda_t to_fp16_cuda = ggml_get_to_fp16_nc_cuda(src1->type);
const cuda_t * src0_ptr = nullptr;
const cuda_t * src1_ptr = nullptr;
ggml_cuda_pool_alloc<cuda_t> src0_alloc(ctx.pool());
ggml_cuda_pool_alloc<cuda_t> src1_alloc(ctx.pool());
// Handle src0
src0_ptr = (const cuda_t *) src0->data;
// Handle src1 - convert if necessary
if (src1->type == src0_type) {
src1_ptr = (const cuda_t *) src1->data;
} else {
// Convert src1 to target type using traits conversion functions
const int64_t ne_src1 = ggml_nelements(src1);
src1_f16_alloc.alloc(ne_src1);
GGML_ASSERT(to_fp16_cuda != nullptr);
src1_alloc.alloc(ne_src1);
to_fp16_cuda(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
src1_f16 = src1_f16_alloc.get();
const auto convert_func = traits::get_nc_converter(src1->type);
GGML_ASSERT(convert_func != nullptr);
convert_func(src1->data, src1_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
src1_ptr = src1_alloc.get();
s11 = ne10;
s12 = ne11*s11;
s13 = ne12*s12;
}
ggml_cuda_pool_alloc<half> dst_f16(ctx.pool());
// Setup destination buffer
ggml_cuda_pool_alloc<cuda_t> dst_temp(ctx.pool());
char * dst_t;
cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F;
cudaDataType_t cu_data_type = CUDA_R_16F;
// dst strides
size_t nbd2 = dst->nb[2];
size_t nbd3 = dst->nb[3];
const half alpha_f16 = 1.0f;
const half beta_f16 = 0.0f;
cublasComputeType_t cu_compute_type = traits::compute_type;
cudaDataType_t cu_data_type = traits::data_type;
cudaDataType_t cu_data_type_a = traits::data_type;
cudaDataType_t cu_data_type_b = traits::data_type;
const void * alpha = traits::get_alpha();
const void * beta = traits::get_beta();
const float alpha_f32 = 1.0f;
const float beta_f32 = 0.0f;
const void * alpha = &alpha_f16;
const void * beta = &beta_f16;
const float beta_f32 = 0.0f;
if (dst->op_params[0] == GGML_PREC_DEFAULT) {
dst_t = (char *) dst_f16.alloc(ne_dst);
nbd2 /= sizeof(float) / sizeof(half);
nbd3 /= sizeof(float) / sizeof(half);
if constexpr (src0_type == GGML_TYPE_F32) {
dst_t = (char *) dst_ddf; // Direct F32 output
} else {
dst_t = (char *) dst_temp.alloc(ne_dst);
nbd2 /= sizeof(float) / sizeof(cuda_t);
nbd3 /= sizeof(float) / sizeof(cuda_t);
}
} else {
dst_t = (char *) dst_ddf;
cu_compute_type = CUBLAS_COMPUTE_32F;
cu_data_type = CUDA_R_32F;
cu_data_type = CUDA_R_32F;
alpha = &alpha_f32;
beta = &beta_f32;
beta = &beta_f32;
}
int id = ggml_cuda_get_device();
@ -1856,7 +1904,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
if (GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
cu_compute_type = CUBLAS_COMPUTE_32F;
alpha = &alpha_f32;
beta = &beta_f32;
beta = &beta_f32;
}
GGML_ASSERT(ne12 % ne02 == 0);
@ -1866,35 +1914,15 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
const int64_t r2 = ne12/ne02;
const int64_t r3 = ne13/ne03;
#if 0
// use cublasGemmEx
{
for (int i13 = 0; i13 < ne13; ++i13) {
for (int i12 = 0; i12 < ne12; ++i12) {
int i03 = i13 / r3;
int i02 = i12 / r2;
CUBLAS_CHECK(
cublasGemmEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
ne01, ne11, ne10,
alpha, (const char *) src0_f16 + i03*nb03 + i02*nb02, CUDA_R_16F, nb01/sizeof(half),
src1_f16 + i13*s13 + i12*s12, CUDA_R_16F, s11,
beta, ( char *) dst_t + i13*nbd3 + i12*nbd2, cu_data_type, ne0,
cu_compute_type,
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
}
}
}
#else
if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
// there is no broadcast and src0, src1 are contiguous across dims 2, 3
// use cublasGemmStridedBatchedEx
CUBLAS_CHECK(
cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
ne01, ne11, ne10,
alpha, src0_f16, CUDA_R_16F, nb01/nb00, nb02/nb00, // strideA
src1_f16, CUDA_R_16F, s11, s12, // strideB
beta, dst_t, cu_data_type, ne0, ne1*ne0, // strideC
alpha, src0_ptr, cu_data_type_a, nb01/nb00, nb02/nb00, // strideA
src1_ptr, cu_data_type_b, s11, s12, // strideB
beta, dst_t, cu_data_type, ne0, ne1*ne0, // strideC
ne12*ne13,
cu_compute_type,
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
@ -1905,34 +1933,55 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
ggml_cuda_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
ggml_cuda_pool_alloc< void *> ptrs_dst(ctx.pool(), 1*ne23);
size_t src1_stride_size = sizeof(cuda_t);
dim3 block_dims(ne13, ne12);
k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
src0_f16, src1_f16, dst_t,
src0_ptr, src1_ptr, dst_t,
ptrs_src.get(), ptrs_dst.get(),
ne12, ne13,
ne23,
nb02, nb03,
src1->type == GGML_TYPE_F16 ? nb12 : s12*sizeof(half),
src1->type == GGML_TYPE_F16 ? nb13 : s13*sizeof(half),
(src1->type == src0_type) ? nb12 : s12*src1_stride_size,
(src1->type == src0_type) ? nb13 : s13*src1_stride_size,
nbd2, nbd3,
r2, r3);
CUDA_CHECK(cudaGetLastError());
CUBLAS_CHECK(
cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
ne01, ne11, ne10,
alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F, nb01/nb00,
(const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F, s11,
beta, ( void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne0,
alpha, (const void **) (ptrs_src.get() + 0*ne23), cu_data_type_a, nb01/nb00,
(const void **) (ptrs_src.get() + 1*ne23), cu_data_type_b, s11,
beta, ( void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne0,
ne23,
cu_compute_type,
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
}
#endif
if (dst->op_params[0] == GGML_PREC_DEFAULT && cu_data_type == CUDA_R_16F) {
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
to_fp32_cuda(dst_f16.get(), dst_ddf, ne_dst, main_stream);
// Convert output back to F32 if needed
if (dst->op_params[0] == GGML_PREC_DEFAULT && cu_data_type != CUDA_R_32F) {
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(traits::ggml_type_val);
to_fp32_cuda(dst_temp.get(), dst_ddf, ne_dst, main_stream);
}
}
static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 || src0->type == GGML_TYPE_F32);
switch (src0->type) {
case GGML_TYPE_F32:
ggml_cuda_mul_mat_batched_cublas_impl<GGML_TYPE_F32>(ctx, src0, src1, dst);
break;
case GGML_TYPE_BF16:
ggml_cuda_mul_mat_batched_cublas_impl<GGML_TYPE_BF16>(ctx, src0, src1, dst);
break;
case GGML_TYPE_F16:
ggml_cuda_mul_mat_batched_cublas_impl<GGML_TYPE_F16>(ctx, src0, src1, dst);
break;
default:
GGML_ABORT("Unsupported type");
}
}
@ -1984,6 +2033,12 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
//printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
//printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
//TODO update for generic tensor parallelism
const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
bool use_batched_cublas_f16 = src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16);
bool use_batched_cublas_bf16 = src0->type == GGML_TYPE_BF16 && bf16_mma_hardware_available(cc);
bool use_batched_cublas_f32 = src0->type == GGML_TYPE_F32;
if (!split && use_mul_mat_vec) {
// the custom F16 vector kernel can be used over batched cuBLAS GEMM
// but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
@ -1992,8 +2047,8 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
} else if (!split && use_mul_mat_q) {
ggml_cuda_mul_mat_q(ctx, src0, src1, nullptr, dst);
} else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) &&
!ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
} else if (!split && (use_batched_cublas_f16 || use_batched_cublas_bf16 || use_batched_cublas_f32)
&& !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
// general KQ + KQV multi-batch without FlashAttention
ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
} else if (use_mul_mat_vec) {
@ -2248,6 +2303,27 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
return false;
}
break;
case GGML_OP_GLU:
switch (ggml_get_glu_op(dst)) {
case GGML_GLU_OP_REGLU:
ggml_cuda_op_reglu(ctx, dst);
break;
case GGML_GLU_OP_GEGLU:
ggml_cuda_op_geglu(ctx, dst);
break;
case GGML_GLU_OP_SWIGLU:
ggml_cuda_op_swiglu(ctx, dst);
break;
case GGML_GLU_OP_GEGLU_ERF:
ggml_cuda_op_geglu_erf(ctx, dst);
break;
case GGML_GLU_OP_GEGLU_QUICK:
ggml_cuda_op_geglu_quick(ctx, dst);
break;
default:
return false;
}
break;
case GGML_OP_NORM:
ggml_cuda_op_norm(ctx, dst);
break;
@ -3041,6 +3117,18 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
return false;
}
break;
case GGML_OP_GLU:
switch (ggml_get_glu_op(op)) {
case GGML_GLU_OP_REGLU:
case GGML_GLU_OP_GEGLU:
case GGML_GLU_OP_SWIGLU:
case GGML_GLU_OP_GEGLU_ERF:
case GGML_GLU_OP_GEGLU_QUICK:
return ggml_is_contiguous_1(op->src[0]);
default:
return false;
}
break;
case GGML_OP_MUL_MAT:
case GGML_OP_MUL_MAT_ID:
{
@ -3112,6 +3200,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
switch (op->src[0]->type) {
case GGML_TYPE_F16:
case GGML_TYPE_F32:
case GGML_TYPE_BF16:
case GGML_TYPE_I32:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
@ -3241,12 +3331,26 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
case GGML_OP_COS:
case GGML_OP_CLAMP:
case GGML_OP_LOG:
case GGML_OP_SSM_SCAN:
case GGML_OP_SSM_CONV:
return true;
case GGML_OP_SSM_SCAN: {
if (op->src[3]->ne[0] == 1) {
// Mamba2
// (kernel only supports d_state == 128 && d_head % 16 == 0)
return op->src[0]->ne[0] == 128 && op->src[0]->ne[1] % 16 == 0;
} else {
// Mamba
// (kernel only supports d_state == 16, d_head == 1, n_head % 128 == 0, n_group == 1)
return op->src[0]->ne[0] == 16 && op->src[0]->ne[1] == 1 && op->src[0]->ne[2] % 128 == 0 && op->src[4]->ne[1] == 1;
}
}
case GGML_OP_SSM_CONV: {
// assumes d_inner % threads == 0
return op->src[0]->ne[1] % 128 == 0;
}
case GGML_OP_CONT:
return op->src[0]->type != GGML_TYPE_BF16;
case GGML_OP_DIAG_MASK_INF:
return true;
case GGML_OP_SOFT_MAX:
return true;
case GGML_OP_SOFT_MAX_BACK: {
@ -3295,6 +3399,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
if (op->src[0]->ne[0] == 192) {
return false;
}
// TODO: support broadcast
// note: this was initially implemented in https://github.com/ggml-org/llama.cpp/pull/14500, but
// the interface of ggml_flash_attn_ext() changed in https://github.com/ggml-org/llama.cpp/pull/14505
if (op->src[0]->ne[3] != 1) {
return false;
}

10
ggml/src/ggml-cuda/mmq.cuh
View File

@ -3016,14 +3016,8 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
const int nbytes_shared = mmq_get_nbytes_shared<type>(mmq_x, mmq_y, cc);
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
if (!shared_memory_limit_raised[id]) {
CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
shared_memory_limit_raised[id] = true;
}
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, mmq_x, MMQ_NWARPS, false>), nbytes_shared);
CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, mmq_x, MMQ_NWARPS, true>), nbytes_shared);
const int nty = (args.nrows_x + mmq_y - 1) / mmq_y;
const int ntx = (args.ncols_dst + mmq_x - 1) / mmq_x;

177
ggml/src/ggml-cuda/softmax.cu
View File

@ -2,6 +2,7 @@
#include "ggml.h"
#include "softmax.cuh"
#include <cstdint>
#include <utility>
template <typename T>
static __device__ __forceinline__ float t2f32(T val) {
@ -13,6 +14,29 @@ __device__ float __forceinline__ t2f32<half>(half val) {
return __half2float(val);
}
struct soft_max_params {
int64_t nheads;
uint32_t n_head_log2;
int64_t ncols;
int64_t nrows_x;
int64_t nrows_y;
int64_t ne00;
int64_t ne01;
int64_t ne02;
int64_t ne03;
int64_t nb11;
int64_t nb12;
int64_t nb13;
int64_t ne12;
int64_t ne13;
float scale;
float max_bias;
float m0;
float m1;
};
// When ncols_template == 0 the bounds for the loops in this function are not known and can't be unrolled.
// As we want to keep pragma unroll for all other cases we supress the clang transformation warning here.
#ifdef __clang__
@ -21,16 +45,24 @@ __device__ float __forceinline__ t2f32<half>(half val) {
#endif // __clang__
template <bool use_shared, int ncols_template, int block_size_template, typename T>
static __global__ void soft_max_f32(
const float * x, const T * mask, float * dst, const int ncols_par, const int nrows_y,
const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
const float * x, const T * mask, float * dst, const soft_max_params p) {
const int ncols = ncols_template == 0 ? p.ncols : ncols_template;
const int tid = threadIdx.x;
const int rowx = blockIdx.x;
const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension
const int64_t i03 = blockIdx.z;
const int64_t i02 = blockIdx.y;
const int64_t i01 = blockIdx.x;
//TODO: noncontigous inputs/outputs
const int rowx = blockIdx.x + blockIdx.y * gridDim.x + blockIdx.z * gridDim.x * gridDim.y;
const int64_t i11 = i01;
const int64_t i12 = i02 % p.ne12;
const int64_t i13 = i03 % p.ne13;
x += int64_t(rowx)*ncols;
mask += int64_t(rowy)*ncols * (mask != nullptr);
mask += (i11*p.nb11 + i12*p.nb12 + i13*p.nb13) / sizeof(T) * (mask != nullptr);
dst += int64_t(rowx)*ncols;
const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
@ -38,7 +70,7 @@ static __global__ void soft_max_f32(
const int warp_id = threadIdx.x / WARP_SIZE;
const int lane_id = threadIdx.x % WARP_SIZE;
const float slope = get_alibi_slope(max_bias, rowx/nrows_y, n_head_log2, m0, m1);
const float slope = get_alibi_slope(p.max_bias, i02, p.n_head_log2, p.m0, p.m1);
extern __shared__ float data_soft_max_f32[];
float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication
@ -55,7 +87,7 @@ static __global__ void soft_max_f32(
break;
}
const float val = x[col]*scale + (mask ? slope*t2f32(mask[col]) : 0.0f);
const float val = x[col]*p.scale + (mask ? slope*t2f32(mask[col]) : 0.0f);
vals[col] = val;
max_val = max(max_val, val);
@ -150,64 +182,58 @@ static __global__ void soft_max_back_f32(
}
}
template<int... Ns, typename T>
static void launch_soft_max_kernels(const float * x, const T * mask, float * dst,
const soft_max_params & p, cudaStream_t stream, dim3 block_dims, dim3 block_nums, size_t nbytes_shared)
{
const int id = ggml_cuda_get_device();
const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
auto launch_kernel = [=](auto I) -> bool {
constexpr int ncols = decltype(I)::value;
constexpr int block = (ncols > 1024 ? 1024 : ncols);
if (p.ncols == ncols) {
CUDA_SET_SHARED_MEMORY_LIMIT((soft_max_f32<true, ncols, block, T>), smpbo);
soft_max_f32<true, ncols, block><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, p);
return true;
}
return false;
};
// unary fold over launch_kernel
if ((launch_kernel(std::integral_constant<int, Ns>{}) || ...)) {
return;
}
//default case
CUDA_SET_SHARED_MEMORY_LIMIT((soft_max_f32<true, 0, 0, T>), smpbo);
soft_max_f32<true, 0, 0><<<block_nums, block_dims, nbytes_shared, stream>>>(x, mask, dst, p);
}
template<typename T>
static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) {
static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const soft_max_params & params, cudaStream_t stream) {
int nth = WARP_SIZE;
const int64_t ncols_x = params.ncols;
while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
const dim3 block_dims(nth, 1, 1);
const dim3 block_nums(nrows_x, 1, 1);
const dim3 block_nums(params.ne01, params.ne02, params.ne03);
const size_t nbytes_shared = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
const uint32_t n_head = nrows_x/nrows_y;
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
const int id = ggml_cuda_get_device();
const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
// FIXME: this limit could be raised by ~2-4x on Ampere or newer
if (nbytes_shared < ggml_cuda_info().devices[ggml_cuda_get_device()].smpb) {
switch (ncols_x) {
case 32:
soft_max_f32<true, 32, 32><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
case 64:
soft_max_f32<true, 64, 64><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
case 128:
soft_max_f32<true, 128, 128><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
case 256:
soft_max_f32<true, 256, 256><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
case 512:
soft_max_f32<true, 512, 512><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
case 1024:
soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
case 2048:
soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
case 4096:
soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
default:
soft_max_f32<true, 0, 0><<<block_nums, block_dims, nbytes_shared, stream>>>
(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
}
if (nbytes_shared <= smpbo) {
launch_soft_max_kernels<32, 64, 128, 256, 512, 1024, 2048, 4096>(x, mask, dst, params, stream, block_dims, block_nums, nbytes_shared);
} else {
const size_t nbytes_shared_low = WARP_SIZE*sizeof(float);
soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, dst, params);
}
}
@ -235,10 +261,11 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
const int64_t ne00 = src0->ne[0];
const int64_t nrows_x = ggml_nrows(src0);
const int64_t nrows_y = src0->ne[1];
const int64_t ne00 = src0->ne[0];
float scale = 1.0f;
float max_bias = 0.0f;
@ -247,10 +274,44 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
const int64_t nb11 = src1 ? src1->nb[1] : 1;
const int64_t nb12 = src1 ? src1->nb[2] : 1;
const int64_t nb13 = src1 ? src1->nb[3] : 1;
const int64_t ne12 = src1 ? src1->ne[2] : 1;
const int64_t ne13 = src1 ? src1->ne[3] : 1;
const uint32_t n_head = src0->ne[2];
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
soft_max_params params = {};
params.nheads = src0->ne[2];
params.n_head_log2 = n_head_log2;
params.ncols = ne00;
params.nrows_x = nrows_x;
params.nrows_y = nrows_y;
params.ne00 = src0->ne[0];
params.ne01 = src0->ne[1];
params.ne02 = src0->ne[2];
params.ne03 = src0->ne[3];
params.nb11 = nb11;
params.nb12 = nb12;
params.nb13 = nb13;
params.ne12 = ne12;
params.ne13 = ne13;
params.scale = scale;
params.max_bias = max_bias;
params.m0 = m0;
params.m1 = m1;
if (use_f16) {
soft_max_f32_cuda(src0_d, (const half *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
soft_max_f32_cuda(src0_d, (const half *) src1_d, dst_d, params, stream);
} else {
soft_max_f32_cuda(src0_d, (const float *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
soft_max_f32_cuda(src0_d, (const float *) src1_d, dst_d, params, stream);
}
}

231
ggml/src/ggml-cuda/ssm-scan.cu
View File

@ -4,16 +4,15 @@ template <size_t splitD, size_t N>
__global__ void __launch_bounds__(splitD, 2)
ssm_scan_f32(const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2,
const float * __restrict__ src3, const float * __restrict__ src4, const float * __restrict__ src5,
const int src0_nb1, const int src0_nb2, const int src1_nb0, const int src1_nb1, const int src1_nb2,
const int src1_nb3, const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
float * __restrict__ dst, const int64_t L) {
GGML_UNUSED(src1_nb0);
GGML_UNUSED(src2_nb0);
const int32_t * __restrict__ src6, float * __restrict__ dst,
const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
const int src2_nb1, const int src2_nb2, const int src3_nb1,
const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
const int64_t s_off, const int64_t d_inner, const int64_t L) {
constexpr int warp_size = ggml_cuda_get_physical_warp_size();
const int bidx = blockIdx.x; // split along B
const int bidy = blockIdx.y; // split along D
const int bidx = blockIdx.x; // split along B (sequences)
const int bidy = blockIdx.y; // split along D (d_inner)
const int tid = threadIdx.x;
const int wid = tid / 32;
const int wtid = tid % 32;
@ -24,23 +23,23 @@ __global__ void __launch_bounds__(splitD, 2)
float * smem_A = smem;
float * smem_s0 = smem_A + splitD * stride_sA;
const float * s0_block = (const float *) ((const char *) src0 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
const float * x_block = (const float *) ((const char *) src1 + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
const float * s0_block = (const float *) ((const char *) src0 + src6[bidx] * src0_nb3 + bidy * splitD * src0_nb2);
const float * x_block = (const float *) ((const char *) src1 + (bidx * src1_nb3) + bidy * splitD * sizeof(float));
const float * dt_block = (const float *) ((const char *) src2 + (bidx * src2_nb2) + bidy * splitD * sizeof(float));
const float * A_block = (const float *) ((const char *) src3 + bidy * splitD * src3_nb1);
const float * B_block = (const float *) ((const char *) src4 + (bidx * src4_nb2));
const float * C_block = (const float *) ((const char *) src5 + (bidx * src5_nb2));
float * y_block = (float *) ((char *) dst + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
float * s_block = (float *) ((char *) dst + src1_nb3 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
const float * B_block = (const float *) ((const char *) src4 + (bidx * src4_nb3));
const float * C_block = (const float *) ((const char *) src5 + (bidx * src5_nb3));
float * y_block = (float *) ((char *) dst + (bidx * d_inner * L * sizeof(float)) + bidy * splitD * sizeof(float));
float * s_block = (float *) ((char *) dst + s_off + bidx * src0_nb3 + bidy * splitD * src0_nb2);
const int stride_s0 = src0_nb1 / sizeof(float);
const int stride_x = src1_nb1 / sizeof(float);
const int stride_s0 = src0_nb2 / sizeof(float);
const int stride_x = src1_nb2 / sizeof(float);
const int stride_dt = src2_nb1 / sizeof(float);
const int stride_A = src3_nb1 / sizeof(float);
const int stride_B = src4_nb1 / sizeof(float);
const int stride_C = src5_nb1 / sizeof(float);
const int stride_B = src4_nb2 / sizeof(float);
const int stride_C = src5_nb2 / sizeof(float);
const int stride_s = stride_s0;
const int stride_y = stride_x;
const int stride_y = d_inner;
// can N not be 16? for example 32?
if (N == 16) {
@ -84,24 +83,156 @@ __global__ void __launch_bounds__(splitD, 2)
}
}
// assumes as many threads as d_state
template <int splitH, int d_state>
__global__ void __launch_bounds__(d_state, 1)
ssm_scan_f32_group(
const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2,
const float * __restrict__ src3, const float * __restrict__ src4, const float * __restrict__ src5,
const int32_t * __restrict__ src6, float * __restrict__ dst,
const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
const int src2_nb1, const int src2_nb2, const int src3_nb1,
const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
const int64_t s_off, const int64_t n_head, const int64_t d_head, const int64_t n_group, const int64_t n_tok) {
const int head_idx = (blockIdx.x * splitH) / d_head;
const int head_off = ((blockIdx.x * splitH) % d_head) * sizeof(float);
const int seq_idx = blockIdx.y;
const int group_off = (head_idx & (n_group - 1)) * d_state * sizeof(float);
const float * s0_block = (const float *) ((const char *) src0 + src6[seq_idx] * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
const float * x_block = (const float *) ((const char *) src1 + (seq_idx * src1_nb3) + blockIdx.x * splitH * sizeof(float));
const float * dt_block = (const float *) ((const char *) src2 + (seq_idx * src2_nb2) + head_idx * sizeof(float));
const float * A_block = (const float *) ((const char *) src3 + head_idx * src3_nb1);
const float * B_block = (const float *) ((const char *) src4 + (seq_idx * src4_nb3) + (group_off));
const float * C_block = (const float *) ((const char *) src5 + (seq_idx * src5_nb3) + (group_off));
float * y_block = dst + (seq_idx * n_tok * n_head * d_head) + blockIdx.x * splitH;
float * s_block = (float *) ((char *) dst + s_off + seq_idx * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
// strides across n_seq_tokens
const int stride_x = src1_nb2 / sizeof(float);
const int stride_dt = src2_nb1 / sizeof(float);
const int stride_B = src4_nb2 / sizeof(float);
const int stride_C = src5_nb2 / sizeof(float);
const int stride_y = n_head * d_head;
float state[splitH];
// for the parallel accumulation
__shared__ float stateC[splitH * d_state];
#pragma unroll
for (int j = 0; j < splitH; j++) {
state[j] = s0_block[j * d_state + threadIdx.x];
}
for (int64_t i = 0; i < n_tok; i++) {
// TODO: only calculate dA and dt_soft_plus once per head instead of every splitH head elements
// TODO: only calculate B and C once per head group
// NOTE: dt_soft_plus, dA and x_dt have the same value across threads here.
float dt_soft_plus = dt_block[i * stride_dt];
if (dt_soft_plus <= 20.0f) {
dt_soft_plus = log1pf(expf(dt_soft_plus));
}
const float dA = expf(dt_soft_plus * A_block[0]);
const float B = B_block[i * stride_B + threadIdx.x];
const float C = C_block[i * stride_C + threadIdx.x];
// across d_head
#pragma unroll
for (int j = 0; j < splitH; j++) {
const float x_dt = x_block[i * stride_x + j] * dt_soft_plus;
state[j] = (state[j] * dA) + (B * x_dt);
stateC[j * d_state + threadIdx.x] = state[j] * C;
}
__syncthreads();
// parallel accumulation for stateC
// TODO: simplify
{
static_assert((d_state & -d_state) == d_state, "the state size has to be a power of 2");
static_assert((splitH & -splitH) == splitH, "splitH has to be a power of 2");
// reduce until w matches the warp size
// TODO: does this work even when the physical warp size is 64?
#pragma unroll
for (int w = d_state; w > WARP_SIZE; w >>= 1) {
// (assuming there are d_state threads)
#pragma unroll
for (int j = 0; j < ((w >> 1) * splitH + d_state - 1) / d_state; j++) {
// TODO: check for bank conflicts
const int k = (threadIdx.x % (w >> 1)) + (d_state * (threadIdx.x / (w >> 1))) + j * d_state * (d_state / (w >> 1));
stateC[k] += stateC[k + (w >> 1)];
}
__syncthreads();
}
static_assert(splitH >= d_state / WARP_SIZE);
#pragma unroll
for (int j = 0; j < splitH / (d_state / WARP_SIZE); j++) {
float y = stateC[(threadIdx.x % WARP_SIZE) + d_state * (threadIdx.x / WARP_SIZE) + j * d_state * (d_state / WARP_SIZE)];
y = warp_reduce_sum(y);
// store the above accumulations
if (threadIdx.x % WARP_SIZE == 0) {
const int k = threadIdx.x / WARP_SIZE + j * (d_state / WARP_SIZE);
y_block[i * stride_y + k] = y;
}
}
}
}
// write back the state
#pragma unroll
for (int j = 0; j < splitH; j++) {
s_block[j * d_state + threadIdx.x] = state[j];
}
}
static void ssm_scan_f32_cuda(const float * src0, const float * src1, const float * src2, const float * src3,
const float * src4, const float * src5, const int src0_nb1, const int src0_nb2,
const int src1_nb0, const int src1_nb1, const int src1_nb2, const int src1_nb3,
const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
float * dst, const int64_t N, const int64_t D, const int64_t L, const int64_t B,
const float * src4, const float * src5, const int32_t * src6, float * dst,
const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3, const int src2_nb1,
const int src2_nb2, const int src3_nb1, const int src4_nb2, const int src4_nb3, const int src5_nb2,
const int src5_nb3, const int64_t s_off, const int64_t d_state, const int64_t head_dim,
const int64_t n_head, const int64_t n_group, const int64_t n_tok, const int64_t n_seq,
cudaStream_t stream) {
const int threads = 128;
// todo: consider D cannot be divided,does this situation exist?
GGML_ASSERT(D % threads == 0);
const dim3 blocks(B, (D + threads - 1) / threads, 1);
const int smem_size = (threads * (N + 1) * 2) * sizeof(float);
if (N == 16) {
ssm_scan_f32<128, 16><<<blocks, threads, smem_size, stream>>>(
src0, src1, src2, src3, src4, src5, src0_nb1, src0_nb2, src1_nb0, src1_nb1, src1_nb2, src1_nb3, src2_nb0,
src2_nb1, src2_nb2, src3_nb1, src4_nb1, src4_nb2, src5_nb1, src5_nb2, dst, L);
// NOTE: if you change conditions here, be sure to update the corresponding supports_op condition!
if (src3_nb1 == sizeof(float)) {
// Mamba-2
if (d_state == 128) {
GGML_ASSERT(d_state % threads == 0);
// NOTE: can be any power of two between 4 and 64
const int splitH = 16;
GGML_ASSERT(head_dim % splitH == 0);
const dim3 blocks((n_head * head_dim + (splitH - 1)) / splitH, n_seq, 1);
ssm_scan_f32_group<16, 128><<<blocks, threads, 0, stream>>>(
src0, src1, src2, src3, src4, src5, src6, dst,
src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok);
} else {
GGML_ABORT("doesn't support d_state!=128.");
}
} else {
GGML_ABORT("doesn't support N!=16.");
// Mamba-1
GGML_ASSERT(n_head % threads == 0);
GGML_ASSERT(head_dim == 1);
GGML_ASSERT(n_group == 1);
const dim3 blocks(n_seq, (n_head + threads - 1) / threads, 1);
const int smem_size = (threads * (d_state + 1) * 2) * sizeof(float);
if (d_state == 16) {
ssm_scan_f32<128, 16><<<blocks, threads, smem_size, stream>>>(
src0, src1, src2, src3, src4, src5, src6, dst,
src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
} else {
GGML_ABORT("doesn't support d_state!=16.");
}
}
}
@ -112,30 +243,25 @@ void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const struct ggml_tensor * src3 = dst->src[3]; // A
const struct ggml_tensor * src4 = dst->src[4]; // B
const struct ggml_tensor * src5 = dst->src[5]; // C
// const int64_t d_state = src0->ne[0];
// const int64_t d_inner = src0->ne[1];
// const int64_t l = src1->ne[1];
// const int64_t b = src0->ne[2];
const struct ggml_tensor * src6 = dst->src[6]; // ids
const int64_t nc = src0->ne[0]; // d_state
const int64_t nr = src0->ne[1]; // d_inner
const int64_t n_t = src1->ne[1]; // number of tokens per sequence
const int64_t n_s = src0->ne[2]; // number of sequences in the batch
const int64_t nr = src0->ne[1]; // head_dim or 1
const int64_t nh = src1->ne[1]; // n_head
const int64_t ng = src4->ne[1]; // n_group
const int64_t n_t = src1->ne[2]; // number of tokens per sequence
const int64_t n_s = src1->ne[3]; // number of sequences in the batch
GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst));
const int64_t s_off = ggml_nelements(src1) * sizeof(float);
GGML_ASSERT(ggml_nelements(src1) + nc*nr*nh*n_s == ggml_nelements(dst));
GGML_ASSERT(src0->nb[0] == sizeof(float));
GGML_ASSERT(src1->nb[0] == sizeof(float));
GGML_ASSERT(src2->nb[0] == sizeof(float));
GGML_ASSERT(src3->nb[0] == sizeof(float));
GGML_ASSERT(src4->nb[0] == sizeof(float));
GGML_ASSERT(src5->nb[0] == sizeof(float));
// required for the dot product between s and C
GGML_ASSERT(src0->nb[1] == src0->ne[0] * sizeof(float));
// required for per-sequence offsets for states
GGML_ASSERT(src0->nb[2] == src0->ne[0] * src0->ne[1] * sizeof(float));
// required to get correct offset for state destination (i.e. src1->nb[3])
GGML_ASSERT(src1->nb[3] == src1->ne[0] * src1->ne[1] * src1->ne[2] * sizeof(float));
GGML_ASSERT(src6->nb[0] == sizeof(int32_t));
const float * src0_d = (const float *) src0->data;
const float * src1_d = (const float *) src1->data;
@ -143,13 +269,16 @@ void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const float * src3_d = (const float *) src3->data;
const float * src4_d = (const float *) src4->data;
const float * src5_d = (const float *) src5->data;
const int32_t * src6_d = (const int32_t *) src6->data;
float * dst_d = (float *) dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src6->type == GGML_TYPE_I32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src0->nb[1], src0->nb[2], src1->nb[0],
src1->nb[1], src1->nb[2], src1->nb[3], src2->nb[0], src2->nb[1], src2->nb[2], src3->nb[1],
src4->nb[1], src4->nb[2], src5->nb[1], src5->nb[2], dst_d, nc, nr, n_t, n_s, stream);
ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src6_d, dst_d,
src0->nb[2], src0->nb[3], src1->nb[2], src1->nb[3], src2->nb[1], src2->nb[2],
src3->nb[1], src4->nb[2], src4->nb[3], src5->nb[2], src5->nb[3],
s_off, nc, nr, nh, ng, n_t, n_s, stream);
}

97
ggml/src/ggml-cuda/unary.cu
View File

@ -196,6 +196,103 @@ void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_unary<op_log>(ctx, dst);
}
/* gated ops */
template <float (*op)(float), typename T>
static __global__ void unary_gated_op_kernel(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1) {
const int64_t i = int64_t(blockDim.x)*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
// perform base op and multiply with gate (either offset in same tensor or a separate one)
const int64_t j0 = (i / n) * o0 + (i % n);
const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
dst[i] = (T)(op((float)x[j0]) * (float)g[j1]);
}
template <float (*op)(float), typename T>
static void unary_gated_cuda(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1, cudaStream_t stream) {
const int64_t num_blocks = (k + CUDA_GLU_BLOCK_SIZE - 1) / CUDA_GLU_BLOCK_SIZE;
unary_gated_op_kernel<op><<<num_blocks, CUDA_GLU_BLOCK_SIZE, 0, stream>>>(x, g, dst, k, n, o0, o1);
}
template <float (*op)(float)>
void ggml_cuda_op_unary_gated(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
void * src0_d = src0->data;
void * src1_d = src1 ? src1->data : src0->data;
const int64_t src0_o = src0->nb[1];
const int64_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
void * dst_d = dst->data;
const int64_t nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(src0->nb[0] == ggml_element_size(src0));
GGML_ASSERT(ggml_is_contiguous(dst));
GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
GGML_ASSERT(src0->type == dst->type);
GGML_ASSERT(dst->ne[0] == nc);
GGML_ASSERT(ggml_nrows(dst) == ggml_nrows(src0));
if (src1) {
GGML_ASSERT(ggml_is_contiguous_1(src1));
GGML_ASSERT(src1->nb[0] == ggml_element_size(src1));
GGML_ASSERT(src1->ne[0] == nc);
GGML_ASSERT(src0->type == src1->type);
}
const int32_t swapped = ((const int32_t *) dst->op_params)[1];
if (src0->type == GGML_TYPE_F16) {
half * src0_p = (half *) src0_d;
half * src1_p = (half *) src1_d;
if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}
unary_gated_cuda<op>(src0_p, src1_p, (half *)dst_d, ggml_nelements(dst), nc, src0_o / sizeof(half), src1_o / sizeof(half), stream);
} else {
float * src0_p = (float *) src0_d;
float * src1_p = (float *) src1_d;
if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}
unary_gated_cuda<op>(src0_p, src1_p, (float *)dst_d, ggml_nelements(dst), nc, src0_o / sizeof(float), src1_o / sizeof(float), stream);
}
}
void ggml_cuda_op_reglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_unary_gated<op_relu>(ctx, dst);
}
void ggml_cuda_op_geglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_unary_gated<op_gelu>(ctx, dst);
}
void ggml_cuda_op_swiglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_unary_gated<op_silu>(ctx, dst);
}
void ggml_cuda_op_geglu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_unary_gated<op_gelu_erf>(ctx, dst);
}
void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_unary_gated<op_gelu_quick>(ctx, dst);
}
/* silu_back */
static __device__ __forceinline__ float op_silu_back(float grad, float x) {

11
ggml/src/ggml-cuda/unary.cuh
View File

@ -15,6 +15,7 @@
#define CUDA_SQRT_BLOCK_SIZE 256
#define CUDA_SIN_BLOCK_SIZE 256
#define CUDA_COS_BLOCK_SIZE 256
#define CUDA_GLU_BLOCK_SIZE 256
void ggml_cuda_op_abs(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
@ -57,3 +58,13 @@ void ggml_cuda_op_sin(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_reglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_geglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_swiglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_geglu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

64
ggml/src/ggml-impl.h
View File

@ -301,6 +301,7 @@ struct ggml_cgraph {
struct ggml_tensor ** grads; // the outputs of these tensors are the gradients of the nodes
struct ggml_tensor ** grad_accs; // accumulators for node gradients
struct ggml_tensor ** leafs; // tensors with constant data
int32_t * use_counts;// number of uses of each tensor, indexed by hash table slot
struct ggml_hash_set visited_hash_set;
@ -467,13 +468,76 @@ static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
// return true if the node's results are only used by N other nodes
// and can be fused into their calculations.
static inline bool ggml_node_has_n_uses(const struct ggml_cgraph * cgraph, int node_idx, int32_t n_uses) {
const struct ggml_tensor * node = cgraph->nodes[node_idx];
// check the use count against how many we're replacing
size_t hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
if (!ggml_bitset_get(cgraph->visited_hash_set.used, hash_pos) || cgraph->use_counts[hash_pos] != n_uses) {
return false;
}
// if node is a view, some other node might be using the intermediate result
// via the view source.
if (node->view_src) {
return false;
}
// If the user requested output for the node, can't fuse
if (node->flags & GGML_TENSOR_FLAG_OUTPUT) {
return false;
}
return true;
}
// Returns true if nodes [i, i+ops.size()) are the sequence of ggml_ops in ops[]
// and are fusable. Nodes are considered fusable according to this function if:
// - all nodes except the last have only one use and are not views/outputs (see ggml_node_has_N_uses).
// - all nodes except the last are a src of the following node.
// - all nodes are the same shape.
// TODO: Consider allowing GGML_OP_NONE nodes in between
static inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, const enum ggml_op * ops, int num_ops) {
if (node_idx + num_ops > cgraph->n_nodes) {
return false;
}
for (int i = 0; i < num_ops; ++i) {
struct ggml_tensor * node = cgraph->nodes[node_idx + i];
if (node->op != ops[i]) {
return false;
}
if (i < num_ops - 1 && !ggml_node_has_n_uses(cgraph, node_idx + i, 1)) {
return false;
}
if (i > 0) {
struct ggml_tensor * prev = cgraph->nodes[node_idx + i - 1];
if (node->src[0] != prev && node->src[1] != prev) {
return false;
}
if (!ggml_are_same_shape(node, prev)) {
return false;
}
}
}
return true;
}
#ifdef __cplusplus
}
#endif
#ifdef __cplusplus
#include <initializer_list>
#include <vector>
// nicer C++ syntax for ggml_can_fuse
inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops) {
return ggml_can_fuse(cgraph, node_idx, ops.begin(), (int)ops.size());
}
// expose GGUF internals for test code
GGML_API size_t gguf_type_size(enum gguf_type type);
GGML_API struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_params params);

166
ggml/src/ggml-kompute/CMakeLists.txt
View File

@ -1,166 +0,0 @@
find_package(Vulkan COMPONENTS glslc REQUIRED)
find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
if (NOT glslc_executable)
message(FATAL_ERROR "glslc not found")
endif()
ggml_add_backend_library(ggml-kompute
ggml-kompute.cpp
../../include/ggml-kompute.h
)
target_link_libraries(ggml-kompute PRIVATE ggml-base kompute)
target_include_directories(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
function(compile_shader)
set(options)
set(oneValueArgs)
set(multiValueArgs SOURCES)
cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
foreach(source ${compile_shader_SOURCES})
get_filename_component(filename ${source} NAME)
set(spv_file ${filename}.spv)
add_custom_command(
OUTPUT ${spv_file}
DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
COMMENT "Compiling ${source} to ${spv_file}"
)
get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
set(FILE_NAME "shader${RAW_FILE_NAME}")
string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
if(CMAKE_GENERATOR MATCHES "Visual Studio")
add_custom_command(
OUTPUT ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
DEPENDS ${spv_file} xxd
COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
)
else()
add_custom_command(
OUTPUT ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
DEPENDS ${spv_file} xxd
COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
)
endif()
endforeach()
endfunction()
if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
message(STATUS "Kompute found")
set(KOMPUTE_OPT_LOG_LEVEL Error CACHE STRING "Kompute log level")
add_subdirectory(kompute)
# Compile our shaders
compile_shader(SOURCES
kompute-shaders/op_scale.comp
kompute-shaders/op_scale_8.comp
kompute-shaders/op_add.comp
kompute-shaders/op_addrow.comp
kompute-shaders/op_mul.comp
kompute-shaders/op_silu.comp
kompute-shaders/op_relu.comp
kompute-shaders/op_gelu.comp
kompute-shaders/op_softmax.comp
kompute-shaders/op_norm.comp
kompute-shaders/op_rmsnorm.comp
kompute-shaders/op_diagmask.comp
kompute-shaders/op_mul_mat_mat_f32.comp
kompute-shaders/op_mul_mat_f16.comp
kompute-shaders/op_mul_mat_q8_0.comp
kompute-shaders/op_mul_mat_q4_0.comp
kompute-shaders/op_mul_mat_q4_1.comp
kompute-shaders/op_mul_mat_q4_k.comp
kompute-shaders/op_mul_mat_q6_k.comp
kompute-shaders/op_getrows_f32.comp
kompute-shaders/op_getrows_f16.comp
kompute-shaders/op_getrows_q4_0.comp
kompute-shaders/op_getrows_q4_1.comp
kompute-shaders/op_getrows_q6_k.comp
kompute-shaders/op_rope_norm_f16.comp
kompute-shaders/op_rope_norm_f32.comp
kompute-shaders/op_rope_neox_f16.comp
kompute-shaders/op_rope_neox_f32.comp
kompute-shaders/op_cpy_f16_f16.comp
kompute-shaders/op_cpy_f16_f32.comp
kompute-shaders/op_cpy_f32_f16.comp
kompute-shaders/op_cpy_f32_f32.comp
)
# Create a custom target for our generated shaders
add_custom_target(generated_shaders DEPENDS
shaderop_scale.h
shaderop_scale_8.h
shaderop_add.h
shaderop_addrow.h
shaderop_mul.h
shaderop_silu.h
shaderop_relu.h
shaderop_gelu.h
shaderop_softmax.h
shaderop_norm.h
shaderop_rmsnorm.h
shaderop_diagmask.h
shaderop_mul_mat_mat_f32.h
shaderop_mul_mat_f16.h
shaderop_mul_mat_q8_0.h
shaderop_mul_mat_q4_0.h
shaderop_mul_mat_q4_1.h
shaderop_mul_mat_q4_k.h
shaderop_mul_mat_q6_k.h
shaderop_getrows_f32.h
shaderop_getrows_f16.h
shaderop_getrows_q4_0.h
shaderop_getrows_q4_1.h
shaderop_getrows_q6_k.h
shaderop_rope_norm_f16.h
shaderop_rope_norm_f32.h
shaderop_rope_neox_f16.h
shaderop_rope_neox_f32.h
shaderop_cpy_f16_f16.h
shaderop_cpy_f16_f32.h
shaderop_cpy_f32_f16.h
shaderop_cpy_f32_f32.h
)
# Create a custom command that depends on the generated_shaders
add_custom_command(
OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
DEPENDS generated_shaders
COMMENT "Ensuring shaders are generated before compiling ggml-kompute.cpp"
)
# Add the stamp to the main sources to ensure dependency tracking
target_sources(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
else()
message(WARNING "Kompute not found")
endif()

2251
ggml/src/ggml-kompute/ggml-kompute.cpp
View File

File diff suppressed because it is too large Load Diff

1
ggml/src/ggml-kompute/kompute

@ -1 +0,0 @@
Subproject commit 4565194ed7c32d1d2efa32ceab4d3c6cae006306

112
ggml/src/ggml-kompute/kompute-shaders/common.comp
View File

@ -1,112 +0,0 @@
#extension GL_EXT_shader_16bit_storage: require
#extension GL_EXT_shader_8bit_storage: require
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
#extension GL_EXT_shader_explicit_arithmetic_types_int64: require
#extension GL_EXT_control_flow_attributes: enable
#extension GL_KHR_shader_subgroup_arithmetic : require
#extension GL_EXT_debug_printf : enable
#define QK4_0 32
#define QK4_1 32
#define GELU_COEF_A 0.044715
#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
#define TWOPI_F 6.283185307179586f
#define QK_K 256
#define K_SCALE_SIZE 12
#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
#define sizeof_block_q4_0 0x12
struct block_q4_0 {
float16_t d;
uint8_t qs[QK4_0 / 2];
};
mat4 dequantize_q4_0(const block_q4_0 xb, uint il) {
const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
const float d2 = d1 / 256.f;
const float md = -8.f * xb.d;
const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
const uint16_t mask1 = mask0 << 8;
mat4 reg;
for (int i=0;i<8;i++) {
uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
reg[i/2][2*(i%2)+0] = d1 * (b & mask0) + md;
reg[i/2][2*(i%2)+1] = d2 * (b & mask1) + md;
}
return reg;
}
#define sizeof_block_q4_1 0x14
struct block_q4_1 {
float16_t d;
float16_t m;
uint8_t qs[QK4_1 / 2];
};
mat4 dequantize_q4_1(const block_q4_1 xb, uint il) {
const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
const float d2 = d1 / 256.f;
const float m = xb.m;
const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
const uint16_t mask1 = mask0 << 8;
mat4 reg;
for (int i=0;i<8;i++) {
uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
reg[i/2][2*(i%2)+0] = ((b & mask0) * d1) + m;
reg[i/2][2*(i%2)+1] = ((b & mask1) * d2) + m;
}
return reg;
}
#define sizeof_block_q4_k 144
struct block_q4_k {
float16_t d;
float16_t dmin;
uint8_t scales[K_SCALE_SIZE];
uint8_t qs[QK_K/2];
};
#define sizeof_block_q6_k 210
struct block_q6_k {
uint8_t ql[QK_K/2]; // quants, lower 4 bits
uint8_t qh[QK_K/4]; // quants, upper 2 bits
int8_t scales[QK_K/16]; // scales, quantized with 8 bits
float16_t d; // super-block scale
};
mat4 dequantize_q6_k(const block_q6_k xb, uint il) {
const float16_t d_all = xb.d;
const uint qlIndex = 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
const uint qhIndex = 32*(il/8) + 16*(il&1);
float16_t sc = xb.scales[(il%2) + 2 * ((il/2))];
il = (il/2) & 3;
const uint16_t kmask1 = il>1 ? uint16_t(il>2 ? 192 : 48) : uint16_t(il>0 ? 12 : 3);
const uint16_t kmask2 = il>1 ? uint8_t(0xF0) : uint8_t(0x0F);
const float16_t coef = il>1 ? float16_t(1.f/16.f) : float16_t(1.f);
const float16_t ml = float16_t(d_all * sc * 32.f);
const float16_t dl = float16_t(d_all * sc * coef);
mat4 reg;
for (int i = 0; i < 16; ++i) {
const float16_t q = (il&1) != 0 ? ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 2))
: ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 4));
reg[i/4][i%4] = dl * q - ml;
}
return reg;
}
#define QK8_0 32
// struct block_q8_0 {
// float16_t d; // delta
// int8_t qs[QK8_0]; // quants
// };
#define sizeof_block_q8_0 34

58
ggml/src/ggml-kompute/kompute-shaders/op_add.comp
View File

@ -1,58 +0,0 @@
#version 450
#include "common.comp"
layout(local_size_x = 1024) in;
layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int nb00;
int nb01;
int nb02;
int nb03;
int ne10;
int ne11;
int ne12;
int ne13;
int nb10;
int nb11;
int nb12;
int nb13;
int ne0;
int nb0;
int nb1;
int nb2;
int nb3;
//int offs; // TODO: needed for GGML_OP_ACC, see metal code
} pcs;
// general-purpose kernel for addition of two tensors
// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
// cons: not very efficient
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const uint i13 = i03 % pcs.ne13;
const uint i12 = i02 % pcs.ne12;
const uint i11 = i01 % pcs.ne11;
int offs = 0; // TMP (see above)
uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + offs) / 4);
uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11 ) / 4);
uint dst_off = uint((i03*pcs.nb3 + i02*pcs.nb2 + i01*pcs.nb1 + offs) / 4);
for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
const uint i10 = i0 % pcs.ne10;
out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] + inB[pcs.inBOff + src1_off + i10];
}
}

25
ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp
View File

@ -1,25 +0,0 @@
#version 450
#include "common.comp"
layout(local_size_x = 1) in;
layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inAOff;
uint inBOff;
uint outOff;
uint row;
} pcs;
void main() {
const uint baseIndex = gl_WorkGroupID.x * 4;
for (uint x = 0; x < 4; x++) {
const uint i = baseIndex + x;
out_[i + pcs.outOff] = inA[i + pcs.inAOff] + inB[(i % pcs.row) + pcs.inBOff];
}
}

52
ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp
View File

@ -1,52 +0,0 @@
#version 450
#include "common.comp"
#define IN_TYPE float16_t
#define IN_TYPE_SIZE 2
#define OUT_TYPE float16_t
#define OUT_TYPE_SIZE 2
layout(local_size_x = 1024) in;
layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
layout (push_constant) uniform parameter {
uint inOff;
uint outOff;
int ne00;
int ne01;
int ne02;
uint nb00;
uint nb01;
uint nb02;
uint nb03;
int ne0;
int ne1;
int ne2;
uint nb0;
uint nb1;
uint nb2;
uint nb3;
} pcs;
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
out_[dst_data+i00] = OUT_TYPE(in_[src]);
}
}

52
ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp
View File

@ -1,52 +0,0 @@
#version 450
#include "common.comp"
#define IN_TYPE float16_t
#define IN_TYPE_SIZE 2
#define OUT_TYPE float
#define OUT_TYPE_SIZE 4
layout(local_size_x = 1024) in;
layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
layout (push_constant) uniform parameter {
uint inOff;
uint outOff;
int ne00;
int ne01;
int ne02;
uint nb00;
uint nb01;
uint nb02;
uint nb03;
int ne0;
int ne1;
int ne2;
uint nb0;
uint nb1;
uint nb2;
uint nb3;
} pcs;
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
out_[dst_data+i00] = OUT_TYPE(in_[src]);
}
}

52
ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp
View File

@ -1,52 +0,0 @@
#version 450
#include "common.comp"
#define IN_TYPE float
#define IN_TYPE_SIZE 4
#define OUT_TYPE float16_t
#define OUT_TYPE_SIZE 2
layout(local_size_x = 1024) in;
layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
layout (push_constant) uniform parameter {
uint inOff;
uint outOff;
int ne00;
int ne01;
int ne02;
uint nb00;
uint nb01;
uint nb02;
uint nb03;
int ne0;
int ne1;
int ne2;
uint nb0;
uint nb1;
uint nb2;
uint nb3;
} pcs;
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
out_[dst_data+i00] = OUT_TYPE(in_[src]);
}
}

52
ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp
View File

@ -1,52 +0,0 @@
#version 450
#include "common.comp"
#define IN_TYPE float
#define IN_TYPE_SIZE 4
#define OUT_TYPE float
#define OUT_TYPE_SIZE 4
layout(local_size_x = 1024) in;
layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
layout (push_constant) uniform parameter {
uint inOff;
uint outOff;
int ne00;
int ne01;
int ne02;
uint nb00;
uint nb01;
uint nb02;
uint nb03;
int ne0;
int ne1;
int ne2;
uint nb0;
uint nb1;
uint nb2;
uint nb3;
} pcs;
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
out_[dst_data+i00] = OUT_TYPE(in_[src]);
}
}

30
ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp
View File

@ -1,30 +0,0 @@
#version 450
#include "common.comp"
layout(local_size_x = 1) in;
layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inOff;
uint outOff;
uint n_past;
int ne00;
int ne01;
} pcs;
void main() {
const uint i02 = gl_WorkGroupID.z;
const uint i01 = gl_WorkGroupID.y;
const uint i00 = gl_WorkGroupID.x;
const uint index = i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00 + i00;
if (i00 > pcs.n_past + i01) {
out_[index + pcs.outOff] = uintBitsToFloat(0xFF800000);
} else {
out_[index + pcs.outOff] = in_[index + pcs.inOff];
}
}

22
ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp
View File

@ -1,22 +0,0 @@
#version 450
#include "common.comp"
layout(local_size_x = 1) in;
layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inOff;
uint outOff;
} pcs;
void main() {
const uint baseIndex = gl_WorkGroupID.x * 8;
for (uint x = 0; x < 8; x++) {
const uint i = baseIndex + x;
const float y = in_[i + pcs.inOff];
out_[i + pcs.outOff] = 0.5*y*(1.0 + tanh(clamp(SQRT_2_OVER_PI*y*(1.0 + GELU_COEF_A*y*y), -15.0, 15.0)));
}
}

17
ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp
View File

@ -1,17 +0,0 @@
void main() {
const uint i = gl_WorkGroupID.x;
const int r = inB[i + pcs.inBOff];
int z = 0;
for (uint ind = gl_LocalInvocationID.x; ind < pcs.ne00/16; ind += gl_WorkGroupSize.x) {
const uint inIndex = (r * pcs.nb01 + pcs.inAOff) + ind/NL * SIZE_OF_BLOCK;
const mat4 result = dequantize_block(inIndex, ind%NL);
for (uint j = 0; j < 4; ++j) {
for (uint k = 0; k < 4; ++k) {
const uint outIndex = i * pcs.nb1/BYTES_FOR_TYPE + pcs.outOff + z;
out_[outIndex] = result[j][k];
++z;
}
}
}
}

31
ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp
View File

@ -1,31 +0,0 @@
#version 450
#include "common.comp"
layout(local_size_x = 1) in;
layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { int inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int nb01;
int nb1;
} pcs;
void dequantize_row_f16(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
for (int j = 0; j < k; j++) {
out_[y + j] = inA[x + j];
}
}
void main() {
const uint i = gl_WorkGroupID.x;
const int r = inB[i + pcs.inBOff];
dequantize_row_f16(r*pcs.nb01/2/*bytes for float16*/ + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
}

31
ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp
View File

@ -1,31 +0,0 @@
#version 450
#include "common.comp"
layout(local_size_x = 1) in;
layout (binding = 0) readonly buffer tensorInA { float inA[]; };
layout (binding = 1) readonly buffer tensorInB { int inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int nb01;
int nb1;
} pcs;
void dequantize_row_f32(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
for (int j = 0; j < k; j++) {
out_[y + j] = inA[x + j];
}
}
void main() {
const uint i = gl_WorkGroupID.x;
const int r = inB[i + pcs.inBOff];
dequantize_row_f32(r*pcs.nb01/4 + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
}

38
ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp
View File

@ -1,38 +0,0 @@
#version 450
#include "common.comp"
#define NL 2
#define BYTES_FOR_TYPE 4 /*bytes for float*/
#define SIZE_OF_BLOCK sizeof_block_q4_0
layout(local_size_x = 1) in;
layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { int inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int nb01;
int nb1;
} pcs;
block_q4_0 get_unaligned_block_q4_0(uint index) {
block_q4_0 fres;
fres.d = u8BufToFloat16(inA, index);
[[unroll]] for (uint it = 0; it != QK4_0 / 2; it++) {
fres.qs[it] = inA[index+2+it];
}
return fres;
}
mat4 dequantize_block(uint index, uint il) {
const block_q4_0 block = get_unaligned_block_q4_0(index);
return dequantize_q4_0(block, il);
}
#include "op_getrows.comp"

39
ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp
View File

@ -1,39 +0,0 @@
#version 450
#include "common.comp"
#define NL 2
#define BYTES_FOR_TYPE 4 /*bytes for float*/
#define SIZE_OF_BLOCK sizeof_block_q4_1
layout(local_size_x = 1) in;
layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { int inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int nb01;
int nb1;
} pcs;
block_q4_1 get_unaligned_block_q4_1(uint index) {
block_q4_1 fres;
fres.d = u8BufToFloat16(inA, index);
fres.m = u8BufToFloat16(inA, index+2);
[[unroll]] for (uint it = 0; it != QK4_1 / 2; it++) {
fres.qs[it] = inA[index+4+it];
}
return fres;
}
mat4 dequantize_block(uint index, uint il) {
const block_q4_1 block = get_unaligned_block_q4_1(index);
return dequantize_q4_1(block, il);
}
#include "op_getrows.comp"

44
ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp
View File

@ -1,44 +0,0 @@
#version 450
#include "common.comp"
#define NL 16
#define BYTES_FOR_TYPE 4 /*bytes for float*/
#define SIZE_OF_BLOCK sizeof_block_q6_k
layout(local_size_x = 1) in;
layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { int inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int nb01;
int nb1;
} pcs;
block_q6_k get_unaligned_block_q6_k(uint index) {
block_q6_k fres;
[[unroll]] for (uint it = 0; it != QK_K / 2; it++) {
fres.ql[it] = inA[index + it];
}
[[unroll]] for (uint it = 0; it != QK_K / 4; it++) {
fres.qh[it] = inA[index + QK_K/2 + it];
}
[[unroll]] for (uint it = 0; it != QK_K / 16; it++) {
fres.scales[it] = int8_t(inA[index + QK_K/2 + QK_K/4 + it]);
}
fres.d = u8BufToFloat16(inA, index + QK_K/2 + QK_K/4 + QK_K/16);
return fres;
}
mat4 dequantize_block(uint index, uint il) {
const block_q6_k block = get_unaligned_block_q6_k(index);
return dequantize_q6_k(block, il);
}
#include "op_getrows.comp"

52
ggml/src/ggml-kompute/kompute-shaders/op_mul.comp
View File

@ -1,52 +0,0 @@
#version 450
#include "common.comp"
layout(local_size_x = 1024) in;
layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int nb00;
int nb01;
int nb02;
int nb03;
int ne10;
int ne11;
int ne12;
int ne13;
int nb10;
int nb11;
int nb12;
int nb13;
int ne0;
int nb0;
int nb1;
int nb2;
int nb3;
} pcs;
void main() {
const uint i03 = gl_WorkGroupID.z;
const uint i02 = gl_WorkGroupID.y;
const uint i01 = gl_WorkGroupID.x;
const uint i13 = i03 % pcs.ne13;
const uint i12 = i02 % pcs.ne12;
const uint i11 = i01 % pcs.ne11;
uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01) / 4);
uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11) / 4);
uint dst_off = uint((i03*pcs.nb3 + i02*pcs.nb2 + i01*pcs.nb1) / 4);
for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
const uint i10 = i0 % pcs.ne10;
out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] * inB[pcs.inBOff + src1_off + i10];
}
}

69
ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp
View File

@ -1,69 +0,0 @@
#version 450
#include "common.comp"
#extension GL_KHR_shader_subgroup_arithmetic : require
layout(local_size_x_id = 0) in;
layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { float inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int ne01;
int ne02;
uint nb00;
uint nb01;
uint nb02;
uint nb03;
int ne10;
int ne11;
int ne12;
uint nb10;
uint nb11;
uint nb12;
uint nb13;
int ne0;
int ne1;
uint r2;
uint r3;
} pcs;
#define N_F16_F32 4
void main() {
const uint r0 = gl_WorkGroupID.x;
const uint rb = gl_WorkGroupID.y*N_F16_F32;
const uint im = gl_WorkGroupID.z;
const uint i12 = im%pcs.ne12;
const uint i13 = im/pcs.ne12;
const uint offset0 = r0*pcs.nb01 + (i12/pcs.r2)*pcs.nb02 + (i13/pcs.r3)*pcs.nb03;
const uint x = offset0 / 2 + pcs.inAOff; // Based from inA
for (uint row = 0; row < N_F16_F32; ++row) {
uint r1 = rb + row;
if (r1 >= pcs.ne11) {
break;
}
const uint y = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
float sumf = 0;
for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
sumf += float(inA[x+i]) * float(inB[y+i]);
}
const float all_sum = subgroupAdd(sumf);
if (subgroupElect()) {
out_[im*pcs.ne1*pcs.ne0 + r1*pcs.ne0 + r0 + pcs.outOff] = all_sum;
}
}
}

51
ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp
View File

@ -1,51 +0,0 @@
#version 450
#include "common.comp"
#extension GL_KHR_shader_subgroup_arithmetic : require
#extension GL_EXT_debug_printf : enable
// device subgroup size
layout (local_size_x_id = 0) in;
layout(binding = 0) readonly buffer tensorInA { float inA[]; };
layout(binding = 1) readonly buffer tensorInB { float inB[]; };
layout(binding = 2) writeonly buffer tensorOut { float out_[]; };
layout(push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int ne01;
int ne02;
int ne11;
int ne12;
uint nb01;
uint nb02;
uint nb11;
uint nb12;
uint nb1;
uint nb2;
}
pcs;
void main() {
uvec3 gid = gl_WorkGroupID;
uint bc_ab = pcs.ne12 > pcs.ne02 ? gid.z / (pcs.ne12 / pcs.ne02) : gid.z;
uint bc_ba = pcs.ne02 > pcs.ne12 ? gid.z / (pcs.ne02 / pcs.ne12) : gid.z;
const uint x = (gid.x*pcs.nb01 + bc_ab*pcs.nb02) / 4 + pcs.inAOff; // Based from inA
const uint y = (gid.y*pcs.nb11 + bc_ba*pcs.nb12) / 4 + pcs.inBOff; // based from inB
float sum = 0.0f;
for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
sum += float(inA[x+i]) * float(inB[y+i]);
}
const float all_sum = subgroupAdd(sum);
if (subgroupElect()) {
out_[gid.z*(pcs.nb2/4) + gid.y*(pcs.nb1/4) + gid.x + pcs.outOff] = all_sum;
}
}

33
ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp
View File

@ -1,33 +0,0 @@
#version 450
#include "common.comp"
#define BLOCKS_IN_QUANT QK4_0
#define SIZE_OF_BLOCK sizeof_block_q4_0
#define N_ROWS 4
#include "op_mul_mv_q_n_pre.comp"
// The q4_0 version of this function
float block_q_n_dot_y(uint block_index, uint yb, uint il) {
vec2 acc = vec2(0.0, 0.0);
const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
float d = float(u8BufToFloat16(inA, index));
float sumy = 0.0f;
for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
const uint16_t b = u8BufToU16(inA, index + 2 + il + i);
const float yl0 = inB[yb + i];
const float yl1 = inB[yb + i + 1];
const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
sumy += yl0 + yl1 + yl8 + yl9;
acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
}
return d * (sumy * -8.f + acc[0] + acc[1]);
}
#include "op_mul_mv_q_n.comp"

35
ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp
View File

@ -1,35 +0,0 @@
#version 450
#include "common.comp"
#define BLOCKS_IN_QUANT QK4_1
#define SIZE_OF_BLOCK sizeof_block_q4_1
#define N_ROWS 4
#include "op_mul_mv_q_n_pre.comp"
// The q4_1 version of this function
float block_q_n_dot_y(uint block_index, uint yb, uint il) {
vec2 acc = vec2(0.0, 0.0);
const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
float d = float(u8BufToFloat16(inA, index));
float m = float(u8BufToFloat16(inA, index+2));
float sumy = 0.0f;
for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
const uint16_t b = u8BufToU16(inA, index + 4 + il + i);
const float yl0 = inB[yb + i];
const float yl1 = inB[yb + i + 1];
const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
sumy += yl0 + yl1 + yl8 + yl9;
acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
}
return d * (acc[0] + acc[1]) + sumy * m;
}
#include "op_mul_mv_q_n.comp"

140
ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp
View File

@ -1,140 +0,0 @@
#version 450
#include "common.comp"
#define N_DST 4
#define SIZE_OF_BLOCK sizeof_block_q4_k
layout(local_size_x = 4) in;
layout(local_size_y = 8) in;
layout(local_size_z = 1) in;
layout (binding = 0) readonly buffer tensorInA { block_q4_k inA[]; };
layout (binding = 1) readonly buffer tensorInB { float inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int ne10;
int ne0;
int ne1;
int ne01;
int ne02;
int ne12;
uint nb01;
uint nb02;
uint nb03;
uint nb11;
uint nb12;
uint nb13;
uint r2;
uint r3;
} pcs;
void main() {
const uint16_t kmask1 = uint16_t(0x3f3f);
const uint16_t kmask2 = uint16_t(0x0f0f);
const uint16_t kmask3 = uint16_t(0xc0c0);
const uint ix = gl_SubgroupInvocationID/8; // 0...3
const uint it = gl_SubgroupInvocationID%8; // 0...7
const uint iq = it/4; // 0 or 1
const uint ir = it%4; // 0...3
const uint nb = pcs.ne00/QK_K;
const uint r0 = gl_WorkGroupID.x;
const uint r1 = gl_WorkGroupID.y;
const uint im = gl_WorkGroupID.z;
const uint first_row = r0 * N_DST;
const uint ib_row = first_row * nb;
const uint i12 = im%pcs.ne12;
const uint i13 = im/pcs.ne12;
const uint offset0 = first_row*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
const uint offset1 = r1*pcs.nb11 + (i12 )*pcs.nb12 + (i13 )*pcs.nb13;
const uint xblk = offset0 + pcs.inAOff;
const uint y = (offset1 / 4) + pcs.inBOff;
float yl[16];
float yh[16];
float sumf[N_DST] = {0.f, 0.f, 0.f, 0.f};
float all_sum = 0.f;
uint y4 = y + ix * QK_K + 64 * iq + 8 * ir;
for (uint ib = ix; ib < nb; ib += 4) {
const uint blk_idx = ib + xblk;
float sumy[4] = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i+0] = inB[y4+i+ 0]; sumy[0] += yl[i+0];
yl[i+8] = inB[y4+i+ 32]; sumy[1] += yl[i+8];
yh[i+0] = inB[y4+i+128]; sumy[2] += yh[i+0];
yh[i+8] = inB[y4+i+160]; sumy[3] += yh[i+8];
}
for (int row = 0; row < N_DST; row++) {
uint row_idx = row * (pcs.nb01 / SIZE_OF_BLOCK);
uint16_t sc_0 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 0);
uint16_t sc_1 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 2);
uint16_t sc_2 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 4);
uint16_t sc_3 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 6);
uint16_t sc_4 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 8);
uint16_t sc16[4];
sc16[0] = sc_0 & kmask1;
sc16[1] = sc_2 & kmask1;
sc16[2] = ((sc_4 >> 0) & kmask2) | ((sc_0 & kmask3) >> 2);
sc16[3] = ((sc_4 >> 4) & kmask2) | ((sc_2 & kmask3) >> 2);
float acc1[4] = {0.f, 0.f, 0.f, 0.f};
float acc2[4] = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
uint16_t q1 = u8BufToU16(inA[blk_idx + row_idx].qs, 32 * iq + 8 * ir + i);
uint16_t q2 = u8BufToU16(inA[blk_idx + row_idx].qs, 64 + 32 * iq + 8 * ir + i);
acc1[0] += yl[i+0] * (q1 & 0x000F);
acc1[1] += yl[i+1] * (q1 & 0x0F00);
acc1[2] += yl[i+8] * (q1 & 0x00F0);
acc1[3] += yl[i+9] * (q1 & 0xF000);
acc2[0] += yh[i+0] * (q2 & 0x000F);
acc2[1] += yh[i+1] * (q2 & 0x0F00);
acc2[2] += yh[i+8] * (q2 & 0x00F0);
acc2[3] += yh[i+9] * (q2 & 0xF000);
}
uint8_t sc8_0 = uint8_t(sc16[0] & 0xFF);
uint8_t sc8_1 = uint8_t(sc16[0] >> 8 );
uint8_t sc8_2 = uint8_t(sc16[1] & 0xFF);
uint8_t sc8_3 = uint8_t(sc16[1] >> 8 );
uint8_t sc8_4 = uint8_t(sc16[2] & 0xFF);
uint8_t sc8_5 = uint8_t(sc16[2] >> 8 );
uint8_t sc8_6 = uint8_t(sc16[3] & 0xFF);
uint8_t sc8_7 = uint8_t(sc16[3] >> 8 );
float dall = float(inA[blk_idx + row_idx].d);
float dmin = float(inA[blk_idx + row_idx].dmin);
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8_0 +
(acc1[2] + 1.f/256.f * acc1[3]) * sc8_1 * 1.f/16.f +
(acc2[0] + 1.f/256.f * acc2[1]) * sc8_4 +
(acc2[2] + 1.f/256.f * acc2[3]) * sc8_5 * 1.f/16.f) -
dmin * (sumy[0] * sc8_2 + sumy[1] * sc8_3 + sumy[2] * sc8_6 + sumy[3] * sc8_7);
}
y4 += 4 * QK_K;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = subgroupAdd(sumf[row]);
if (subgroupElect()) {
out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = all_sum;
}
}
}

106
ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp
View File

@ -1,106 +0,0 @@
#version 450
#include "common.comp"
#define SIZE_OF_BLOCK sizeof_block_q6_k
layout(local_size_x_id = 0) in;
layout(local_size_y_id = 1) in;
layout(local_size_z = 1) in;
layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { float inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int ne10;
int ne0;
int ne1;
int ne01;
int ne02;
int ne12;
uint nb01;
uint nb02;
uint nb03;
uint nb11;
uint nb12;
uint nb13;
uint r2;
uint r3;
} pcs;
void main() {
const uint8_t kmask1 = uint8_t(0x03);
const uint8_t kmask2 = uint8_t(0x0C);
const uint8_t kmask3 = uint8_t(0x30);
const uint8_t kmask4 = uint8_t(0xC0);
const uint nb = pcs.ne00/QK_K;
const uint r0 = gl_WorkGroupID.x;
const uint r1 = gl_WorkGroupID.y;
const uint im = gl_WorkGroupID.z;
const uint row = (r0 * gl_NumSubgroups + gl_SubgroupID);
const uint i12 = im%pcs.ne12;
const uint i13 = im/pcs.ne12;
const uint x = row*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
const uint yy = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
float sumf = 0;
// bits of invocation ID for gl_SubgroupSize=32:
// x x x x x
// 4 3 2 1 0
// ( tid ) ix
// ip ( il )
const uint block_stride = gl_SubgroupSize / 16; // number of blocks each subgroup processes
const uint tid = gl_SubgroupInvocationID/block_stride; // first block_stride groups have tid=0
const uint ix = gl_SubgroupInvocationID%block_stride; // first block is 0..block_stride-1
const uint ip = tid/8; // first or second half of block (0 or 1)
const uint il = tid%8; // each half has 8 parts, one per scale
const uint n = 4; // 4 scales at a time (and 4 sums)
const uint l0 = n*il; // offset into half-block, 0..28
const uint is = 8*ip + l0/16; // 0, 1, 8, 9
const uint y_offset = 128*ip + l0;
const uint q_offset_l = 64*ip + l0;
const uint q_offset_h = 32*ip + l0;
for (uint i = ix; i < nb; i += block_stride) {
const uint baseIndex = (x + i) * SIZE_OF_BLOCK + pcs.inAOff;
const uint qlIndex = q_offset_l;
const uint q2Index = qlIndex + QK_K/8;
const uint qhIndex = q_offset_h;
const uint y = yy + i * QK_K + y_offset;
float sums[4] = {0.0f, 0.0f, 0.0f, 0.0f};
for (uint l = 0; l < n; ++l) {
const uint8_t currentQ1 = inA[baseIndex + qlIndex + l];
const uint8_t currentQ2 = inA[baseIndex + q2Index + l];
const uint8_t currentQh = inA[baseIndex + QK_K/2 + qhIndex + l];
sums[0] += inB[y+l+ 0] * (int8_t((currentQ1 & 0xF) | ((currentQh & kmask1) << 4)) - 32);
sums[1] += inB[y+l+32] * (int8_t((currentQ2 & 0xF) | ((currentQh & kmask2) << 2)) - 32);
sums[2] += inB[y+l+64] * (int8_t((currentQ1 >> 4) | ((currentQh & kmask3) << 0)) - 32);
sums[3] += inB[y+l+96] * (int8_t((currentQ2 >> 4) | ((currentQh & kmask4) >> 2)) - 32);
}
float d = u8BufToFloat16(inA, baseIndex + QK_K/2 + QK_K/4 + QK_K/16);
sumf += d * (sums[0] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + is]) + sums[1] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 2 + is]) + sums[2] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 4 + is]) + sums[3] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 6 + is]));
}
const float tot = subgroupAdd(sumf);
if (subgroupElect()) {
out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + row + pcs.outOff] = tot;
}
}

73
ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp
View File

@ -1,73 +0,0 @@
#version 450
#include "common.comp"
#include "op_mul_mv_q_n_pre.comp"
#define SIZE_OF_D 2
#define N_DST 4 // each SIMD group works on 4 rows
#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
#define NB_Q8_0 8
void main() {
// NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
if (gl_SubgroupInvocationID > 31)
return;
const int nr = N_DST;
const int nsg = N_SIMDGROUP;
const int nw = N_SIMDWIDTH;
const int nb = pcs.ne00/QK8_0;
const uint r0 = gl_WorkGroupID.x;
const uint r1 = gl_WorkGroupID.y;
const uint im = gl_WorkGroupID.z;
const uint first_row = (r0 * nsg + gl_SubgroupID) * nr;
const uint i12 = im%pcs.ne12;
const uint i13 = im/pcs.ne12;
const uint offset0 = first_row * nb + (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
const uint x = offset0*sizeof_block_q8_0 + pcs.inAOff; // Based from inA
const uint y = r1*pcs.ne10 + im*pcs.ne00*pcs.ne1 + pcs.inBOff; // based from inB
float yl[NB_Q8_0];
float sumf[N_DST]={0.f, 0.f, 0.f, 0.f};
const uint ix = gl_SubgroupInvocationID.x/4;
const uint il = gl_SubgroupInvocationID.x%4;
uint yb = y + ix * QK8_0 + NB_Q8_0*il;
// each thread in a SIMD group deals with NB_Q8_0 quants at a time
for (uint ib = ix; ib < nb; ib += nw/4) {
for (int i = 0; i < NB_Q8_0; ++i) {
yl[i] = inB[yb + i];
}
for (int row = 0; row < nr; row++) {
const uint block_offset = (ib+row*nb) * sizeof_block_q8_0;
float sumq = 0.f;
for (int iq = 0; iq < NB_Q8_0; ++iq) {
const int8_t qs_iq = int8_t(inA[x + block_offset + SIZE_OF_D + NB_Q8_0*il + iq]);
sumq += qs_iq * yl[iq];
}
const float16_t d = u8BufToFloat16(inA, x + block_offset);
sumf[row] += sumq*d;
}
yb += NB_Q8_0 * nw;
}
for (int row = 0; row < nr; ++row) {
const float tot = subgroupAdd(sumf[row]);
if (subgroupElect() && first_row + row < pcs.ne01) {
out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row] = tot;
}
}
}

52
ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp
View File

@ -1,52 +0,0 @@
void main() {
// NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
if (gl_SubgroupInvocationID > 31)
return;
const uint nb = uint(pcs.ne00/BLOCKS_IN_QUANT);
const uint r0 = gl_WorkGroupID.x;
const uint r1 = gl_WorkGroupID.y;
const uint im = gl_WorkGroupID.z;
const uint first_row = (r0 * gl_NumSubgroups + gl_SubgroupID) * N_ROWS;
const uint i12 = im%pcs.ne12;
const uint i13 = im/pcs.ne12;
// pointers to src0 rows
uint ax[N_ROWS];
for (int row = 0; row < N_ROWS; ++row) {
const uint offset0 = (first_row + row)*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
ax[row] = offset0 + pcs.inAOff;
}
const uint y = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
float sumf[N_ROWS] = {0.0f, 0.0f, 0.0f, 0.0f};
const uint ix = gl_SubgroupInvocationID/2;
const uint il = (BLOCKS_IN_QUANT/4)*(gl_SubgroupInvocationID%2);
uint yb = y + ix * BLOCKS_IN_QUANT + il;
//debugPrintfEXT("gl_NumSubgroups=%d, gl_SubgroupID=%d, gl_SubgroupInvocationID=%d, glSubgroupSize=%d, gl_WorkGroupSize.x=%d, gl_WorkGroupSize.y=%d, gl_WorkGroupSize.z=%d\n",
// gl_NumSubgroups, gl_SubgroupID, gl_SubgroupInvocationID, gl_SubgroupSize,
// gl_WorkGroupSize.x, gl_WorkGroupSize.y, gl_WorkGroupSize.z);
for (uint ib = ix; ib < nb; ib += 16) {
for (int row = 0; row < N_ROWS; row++) {
sumf[row] += block_q_n_dot_y(ax[row] + ib, yb, il);
}
yb += BLOCKS_IN_QUANT * 16;
}
for (int row = 0; row < N_ROWS; ++row) {
const float tot = subgroupAdd(sumf[row]);
if (first_row + row < pcs.ne01 && subgroupElect()) {
out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = tot;
}
}
}

28
ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp
View File

@ -1,28 +0,0 @@
layout(local_size_x_id = 0) in;
layout(local_size_y = 8) in;
layout(local_size_z = 1) in;
layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { float inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
layout (push_constant) uniform parameter {
uint inAOff;
uint inBOff;
uint outOff;
int ne00;
int ne01;
int ne02;
int ne10;
int ne12;
int ne0;
int ne1;
uint nb01;
uint nb02;
uint nb03;
uint nb11;
uint nb12;
uint nb13;
uint r2;
uint r3;
} pcs;

84
ggml/src/ggml-kompute/kompute-shaders/op_norm.comp
View File

@ -1,84 +0,0 @@
#version 450
#include "common.comp"
layout(local_size_x = 256) in;
layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
layout(binding = 1) buffer restrict tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inOff;
uint outOff;
uint ne00;
uint nb01;
float eps;
} pcs;
shared float sum[gl_WorkGroupSize.x];
void main() {
const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
// MEAN
// parallel sum
sum[gl_LocalInvocationID.x] = 0.0;
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
sum[gl_LocalInvocationID.x] += in_[x+i00];
}
// reduce
barrier();
memoryBarrierShared();
[[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
if (gl_LocalInvocationID.x < i) {
sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
}
barrier();
memoryBarrierShared();
}
// broadcast
if (gl_LocalInvocationID.x == 0) {
sum[0] /= float(pcs.ne00);
}
barrier();
memoryBarrierShared();
const float mean = sum[0];
// recenter
const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
out_[y+i00] = in_[x+i00] - mean;
}
// VARIANCE
// parallel sum
sum[gl_LocalInvocationID.x] = 0.0;
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
sum[gl_LocalInvocationID.x] += out_[y+i00] * out_[y+i00];
}
// reduce
barrier();
memoryBarrierShared();
[[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
if (gl_LocalInvocationID.x < i) {
sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
}
barrier();
memoryBarrierShared();
}
// broadcast
if (gl_LocalInvocationID.x == 0) {
sum[0] /= float(pcs.ne00);
}
barrier();
memoryBarrierShared();
const float variance = sum[0];
const float scale = 1.0f/sqrt(variance + pcs.eps);
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
out_[y+i00] *= scale;
}
}

21
ggml/src/ggml-kompute/kompute-shaders/op_relu.comp
View File

@ -1,21 +0,0 @@
#version 450
#include "common.comp"
layout(local_size_x = 1) in;
layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inOff;
uint outOff;
} pcs;
void main() {
const uint baseIndex = gl_WorkGroupID.x * 4;
for (uint x = 0; x < 4; x++) {
const uint i = baseIndex + x;
out_[i + pcs.outOff] = max(0.0, in_[i + pcs.inOff]);
}
}

53
ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp
View File

@ -1,53 +0,0 @@
#version 450
#include "common.comp"
layout(local_size_x = 512) in;
layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
layout(binding = 1) buffer restrict tensorOut { float out_[]; };
layout(push_constant) uniform PushConstants {
uint inOff;
uint outOff;
uint ne00;
uint nb01;
float eps;
} pcs;
shared float sum[gl_WorkGroupSize.x];
void main() {
const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
// parallel sum
sum[gl_LocalInvocationID.x] = 0.0;
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
sum[gl_LocalInvocationID.x] += in_[x+i00] * in_[x+i00];
}
// reduce
barrier();
memoryBarrierShared();
[[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
if (gl_LocalInvocationID.x < i) {
sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
}
barrier();
memoryBarrierShared();
}
// broadcast
if (gl_LocalInvocationID.x == 0) {
sum[0] /= float(pcs.ne00);
}
barrier();
memoryBarrierShared();
const float scale = 1.0f/sqrt(sum[0] + pcs.eps);
const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
out_[y+i00] = in_[x+i00] * scale;
}
}

52
ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp
View File

@ -1,52 +0,0 @@
#version 450
#include "rope_common.comp"
layout(binding = 0) buffer restrict readonly tensorInA { float16_t inA[]; };
layout(binding = 1) buffer restrict readonly tensorInB { int inB[]; };
layout(binding = 2) buffer restrict readonly tensorInC { float inC[]; };
layout(binding = 3) buffer restrict writeonly tensorOut { float16_t out_[]; };
void main() {
const uint i3 = gl_WorkGroupID.z;
const uint i2 = gl_WorkGroupID.y;
const uint i1 = gl_WorkGroupID.x;
float corr_dims[2];
rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
float theta_base = float(inB[pcs.inBOff + i2]);
float inv_ndims = -1.f/pcs.n_dims;
float cos_theta;
float sin_theta;
for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
if (i0 < pcs.n_dims) {
uint ic = i0/2;
float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + ic*pcs.nb00) / 2) + pcs.inAOff; // Based from in
const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + ic*pcs.nb0) / 2) + pcs.outOff; // Based from out_
const float x0 = float(inA[src]);
const float x1 = float(inA[src+pcs.n_dims/2]);
out_[dst_data] = float16_t(x0*cos_theta - x1*sin_theta);
out_[dst_data+pcs.n_dims/2] = float16_t(x0*sin_theta + x1*cos_theta);
} else {
const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 2) + pcs.outOff; // Based from out_
out_[dst_data] = inA[src];
out_[dst_data+1] = inA[src+1];
}
}
}

Some files were not shown because too many files have changed in this diff Show More